Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement filed:
CHOI, Sun et al. Big Data/Artificial Intelligence (Al)-based Drug Discovery Platform and Candidate Discovery. Ewha University - Industry Collaboration Foundation, Ministry of Science and ICT). January 13, 2023, pages 5-61, Retrieved from the Internet: - Unable to determine the reference cited in the IDS as the document does not have a title or anything to identify the cited reference.
The issues noted above fails to comply with 37 CFR 1.98(a)(2), which requires a legible copy of each cited foreign patent document; each non-patent literature publication or that portion which caused it to be listed; and all other information or that portion which caused it to be listed. It has been placed in the application file, but the information referred to therein has not been considered.
Claim Rejections - 35 USC § 112
Claim 2-5, 7-10 rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Regarding claim 2 and analogous claim 7, the terms “a characteristic vector” and “a latent vector” are indefinite as both claim 2 and 7 depend on claim 1 and analogous 6 respectively. Claim 1 and claim 6 both recite “a characteristic vector” and “a latent vector” making claim 2 and 7 indefinite as to what a characteristic and latent vector are referred to or if they are the same as mention in parent claims. All dependent claims inherit the issue.
Regarding claim 3 and analogous claim 8, the terms “a characteristic vector”, “a latent vector”, “an experimental condition vector”, and “ a simulated characteristic vector” are indefinite as both claim 3 and 8 depend on claim 2 and analogous 7 respectively. Parent claims recite “a characteristic vector”, “a latent vector”, “an experimental condition vector”, and “ a simulated characteristic vector” making claim 3 and 8 indefinite as to what a characteristic, latent vector, experimental condition, and simulated characteristic are or if they are the same as mention in parent claims.
Regarding claim 4 and analogous claim 9, the terms “a characteristic vector”, “a latent vector”, “a second regressor”,“ a production condition vector”, and “a weight” are indefinite as both claim 4 and 9 depend on claim 2 and analogous 7 respectively. Parent claims recite “a characteristic vector”, “a latent vector”, “a second regressor”, and “ a production condition vector” making claim 4 and 9 indefinite as to what a characteristic vector, a latent vector, a second regressor, a production condition vector are or if they are the same as mention in parent claims. Further claim 4 recites ““a weight” twice making it indefinite if it refers to the same weight.
Regarding claim 5 and analogous claim 10, the terms “a characteristic vector”, “a latent vector”, “a second regressor”,“ a production condition vector”, and “a weight” are indefinite as both claim 5 and 10 depend on claim 2 and analogous 7 respectively. Parent claims recite “a characteristic vector”, “a latent vector”, “a second regressor”, and “ a production condition vector” making claim 5 and 10 indefinite as to what a characteristic vector, a latent vector, a second regressor, a production condition vector are or if they are the same as mention in parent claims. Further claim 5 recites ““a weight” twice making it indefinite if it refers to the same weight.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 rejected under 35 U.S.C. 101 because the claimed invention is directed to abstract idea without significantly more. The claim(s) recite(s) significantly more. The subject matter eligibility test for products and process is describe below for claim 1 in view of dependent claims.
Regarding claim 1:
Step 1: Is the claim to a process machine manufacture or composition of matter?
Yes – Claim 1 recites a method, which is a method that falls under the statutory categories.
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“[in case that, upon producing a composite using multiple materials, there is an encoder trained to] calculate a latent vector from a characteristic vector indicating a target composite characteristic” - The limitations of claim 1 recites a mathematical process of calculating a latent vector from a characteristic vector (see MPEP 2106.04(a)(2)I).
“calculating, [through a first regressor], a simulated latent vector simulating the latent vector from the input production condition vector;”- The limitations of claim 1 recites a mathematical process of calculating a simulated latent vector (see MPEP 2106.04(a)(2)I).
“[by the inference unit,] calculating, [through a decoder], a simulated characteristic vector [simulating the characteristic vector from the simulated latent vector and an experimental condition vector indicating an experimental condition for measuring the composite characteristic].” - The limitations of claim 1 recites a mathematical process of calculating a simulated characteristic vector (see MPEP 2106.04(a)(2)I).
Step 2 Prong 2: Does the claim recite additional elements that integrate the judicial exception into a particular application? No –
The claim includes the additional element(s):
“A method for mutually inferring a composite characteristic and a composite production condition, the method comprising: in case that, upon producing a composite using multiple materials, there is an encoder trained to [calculate a latent vector from a characteristic vector indicating a target composite characteristic,]”
The additional elements fall under “apply it” as using a generic computer to implement an encoder to calculate the latent vector. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
“when a production condition vector indicating a production condition for obtaining the composite characteristic is received, by an inference unit, [calculating], through a first regressor,”
The additional elements fall under “apply it” as using a generic computer to use an inference unit to calculate though a first regressor. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
“by the inference unit, [calculating], through a decoder, [a simulated characteristic vector] simulating the characteristic vector from the simulated latent vector and an experimental condition vector indicating an experimental condition for measuring the composite characteristic.”
The additional elements fall under “apply it” as using a generic computer to use an inference unit to implement a decoder and calculate the simulated characteristic vector (see MPEP 2106.05(f)).
Step 2B: Does the claim recite additional elements that amount to significantly more than the judicial exception?
No - The claim does not include additional elements that are sufficient to amount to a significantly more than the judicial exemption. As an order whole, the claim is directed to towards calculating vectors for production conditions. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements of indicating and generating a simulated vector fall under using generic computer to apply an exemption. The method does not improve on the function of a computer, transforms an article into another article, nor is it applied by a particular machine, making the claim not patent eligible.
Regarding claim 2:
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“[when a characteristic vector is received, by the inference unit,] calculating a latent vector from the characteristic vector [through the encoder;]” - The limitations recites a mathematical process of calculating a latent vector from a characteristic vector (see MPEP 2106.04(a)(2)I).
“[by the inference unit,] calculating, [through a second regressor,] a simulated production condition vector simulating the production condition vector from the latent vector.” - The limitations recites a mathematical process of calculating a simulated production condition vector (see MPEP 2106.04(a)(2)I).
Step 2A Prong 2, Step 2B: The additional element(s):
“when a characteristic vector is received, by the inference unit, [calculating a latent vector from the characteristic vector] through the encoder;”
The additional elements fall under “apply it” as using a generic computer to implement an encoder to calculate the latent vector (see MPEP 2106.05(f)).
“by the inference unit, [calculating,] through a second regressor, [a simulated production condition vector simulating the production condition vector from the latent vector].”
The additional elements fall under “apply it” as using a generic computer to implement an second regressor to calculate simulated production condition vector (see MPEP 2106.05(f)).
Regarding claim 3:
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“[when a learning unit receives a characteristic vector for learning into the encoder, by the encoder,] calculating a latent vector for learning from the characteristic vector for learning;” - The limitations recites a mathematical process of calculating a latent vector from a characteristic vector (see MPEP 2106.04(a)(2)I).
“[when the learning unit receives a latent vector for learning and an experimental condition vector for learning into the decoder, by the decoder,] calculating a simulated characteristic vector for learning from the latent vector for learning and the experimental condition vector for learning;” - The limitations recites a mathematical process of calculating a simulated characteristic vector for learning (see MPEP 2106.04(a)(2)I).
“[by the learning unit,] calculating a loss representing a difference between the simulated characteristic vector for learning and the characteristic vector for learning;” The limitations recites a mental process of calculating a loss (see MPEP 2106.04(a)(2)III).
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 2, further comprising: before calculating the simulated latent vector,”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
“when a learning unit receives a characteristic vector for learning into the encoder, by the encoder, [calculating a latent vector for learning from the characteristic vector for learning;]”
The additional elements fall under “apply it” as using a generic computer to implement an encoder to calculate the a latent vector for learning (see MPEP 2106.05(f)).
“when the learning unit receives a latent vector for learning and an experimental condition vector for learning into the decoder, by the decoder, [calculating a simulated characteristic vector for learning from the latent vector for learning and the experimental condition vector for learning;]”
The additional elements fall under “apply it” as using a generic computer to implement a decoder to calculating a simulated characteristic vector for learning (see MPEP 2106.05(f)).
“by the learning unit, [calculating a loss representing a difference between the simulated characteristic vector for learning and the characteristic vector for learning;]”
The additional elements fall under “apply it” as using a generic computer to implement to calculate a loss using a learning unit (see MPEP 2106.05(f)).
“by the learning unit, performing optimization to update parameters of the encoder and the decoder to minimize the loss.”
The additional elements fall under “apply it” as using a generic computer to perform optimization on the encoder and decoder (see MPEP 2106.05(f)).
Regarding claim 4:
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“completing the second regressor by deriving a weight for the latent vector for learning using the learning data.” - The limitations recites a mental process of deriving a weight (see MPEP 2106.04(a)(2)III).
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 2, further comprising: before calculating the simulated latent vector,
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
“by a learning unit, preparing a plurality of learning data, each of which includes a latent vector for learning derived from a characteristic vector for learning by the encoder and a production condition vector for learning corresponding to the characteristic vector for learning;”
The additional elements fall under “apply it” as using a generic computer to prepare data using a plurality of learning data (see MPEP 2106.05(f)).
“by the learning unit, preparing a prototype of a second regressor in which the latent vector for learning is used as an input, the production condition vector for learning is used as an output,”
The additional elements fall under “apply it” as using a generic computer to prepare a prototype of a second regressor (see MPEP 2106.05(f)).
“and a weight for the latent vector for learning is not determined”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 5:
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“completing the first regressor by deriving a weight of the production condition vector for learning using the learning data.” - The limitations recites a mental process of deriving a weight (see MPEP 2106.04(a)(2)III).
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 2, further comprising: before calculating the simulated latent vector,
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
“by a learning unit, preparing a plurality of learning data, each of which includes a production condition vector for learning and a latent vector for learning derived from a characteristic vector for learning corresponding to the production condition vector for learning by the encoder;”
The additional elements fall under “apply it” as using a generic computer to prepare data using a plurality of learning data (see MPEP 2106.05(f)).
“by the learning unit, preparing a prototype of a first regressor in which the production condition vector for learning is used an input, the latent vector for learning is used as an output, ”
The additional elements fall under “apply it” as using a generic computer to prepare a prototype of a first regressor (see MPEP 2106.05(f)).
“and a weight of the production condition vector for learning is not determined;”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 6,
Step 1: Is the claim to a process machine manufacture or composition of matter?
No – Claim 6 recites a device, however the claim fails to have the hardware necessary to execute the instructions.
Claim 6 is rejected under 35 U.S.C. 101 because the claimed invention is directed to non-statutory subject matter. The claim does not fall within at least one of the four categories of patent eligible subject matter because it does not describe the system as including any hardware such as processor and memory, therefore, may constitute software per se, which is not a statutory category.
Examiner Note: The Alice/Mayo test is the same regarding the remainder of Step 2A prong 1 and 2 and Step 2B as explained below.
Claims 6-10 recite a system and are analogous to the method of claims 1-5. Therefore, the rejections of claim 1-5 above applies to claims 6-10.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-5 and 6 are rejected under 35 U.S.C. 103 as being unpatentable over Burns et al. (US10059061B2) (“Burns”) in view of Xue, Tianju, et al. "Machine learning generative models for automatic design of multi-material 3D printed composite solids." Extreme Mechanics Letters 41 (2020): 100992 (“Xue”) and Kawamata, Ryota, Shinji Wakao, and Noboru Murata. "Application of conditional variational auto-encoder to magnetic circuit design with magnetic field computation." 2019 22nd International Conference on the Computation of Electromagnetic Fields (COMPUMAG). IEEE, 2019 (“Kawamata”).
Regarding claim 1 and analogous claim 6, as best understood given the 112(b) issue identified above.
Burns teaches A method for mutually inferring a composite characteristic and a composite production condition , the method comprising:
in case that, upon producing a composite using multiple materials, [there is an encoder trained to calculate a latent vector from a characteristic vector indicating a target composite characteristic] (Burns FIG. 2A,
PNG
media_image1.png
606
757
media_image1.png
Greyscale
Col 10 line 21-23, The system 200 may operate during resin infusion and cure to maintain a desired shape of the front of the flow of resin. In yet another example , the system may operate to achieve a desired shape of the composite structure, such as its nominal engineering shape.)
Col 11 line 1-11, The computing device 208 may be coupled to the plurality of sensors 206 and configured to generate control data to achieve a defined quality goal for the component element or composite structure 202. The control data may be generated in accordance with a number of feedback control mechanisms. In some examples, the computing device may be configured to apply the sensor data to a machine-learning algorithm 218 (e.g., computer-readable program code) configured to generate control data to achieve the defined quality goal for the component element or composite structure
Col 12 line 6-12, In some examples, the component element affected by the thermal zones 216 is the resin, either from a pre-impregnated bed of fibers (ply), or bed of fibers that is infused with resin. In these examples, the heating/cooling devices 214 and thereby the thermal zones may be independently controllable to locally heat, cool or maintain the temperature of the resin. [in case that, upon producing a composite using multiple materials,]
Col 12 line 11-23, As explained above, in some examples, manufacture of the composite structure 202 includes resin infusion, which may involve a flow of resin across the bed of fibers to thereby infuse the bed of fibers with resin. In these examples, the defined quality goal for the component element or composite structure may include a desired shape of a front of the flow of resin. The machine-learning algorithm 218 may therefore be configured to generate control data to achieve the desired shape of the front of the flow of resin [A method for mutually inferring a composite characteristic and a composite production condition, the method comprising]),
However Burns does not teach [in case that, upon producing a composite using multiple materials], there is an encoder trained to calculate a latent vector from a characteristic vector indicating a target composite characteristic,
when a production condition vector indicating a production condition for obtaining the composite characteristic is received, by an inference unit, calculating, through a first regressor, a simulated latent vector simulating the latent vector from the input production condition vector;
and by the inference unit, calculating, through a decoder, a simulated characteristic vector simulating the characteristic vector from the simulated latent vector and an experimental condition vector indicating an experimental condition for measuring the composite characteristic.
Xue teaches [in case that, upon producing a composite using multiple materials], there is an encoder trained to calculate a latent vector from a characteristic vector indicating a target composite characteristic (Page 2 In this work, the composite solids are made from two base materials: a commercial hard polyurethane (RPU) with measured Young’s modulus Eh = 1300 MPa and Poisson’s ratio νh = 0.23 and a custom soft silicone (SilDN) [28] with Es = 0.12 MPa andνs = 0.33. We make plane stress assumption throughout the work. We further restrict the composite solids to preserving cubic symmetry as shown in Fig. 1. It can be shown that under these assumptions the constitutive tensor C are fully described by three independent elastic moduli [29,30]. For example, it is common to pick Young’s modulus E, shear modulus G and Poisson’s ratio ν as a complete set of description, in which case the macroscopic stress–strain relationship simplifies to
Page 3, Figure 1,
PNG
media_image2.png
675
1214
media_image2.png
Greyscale
3.1. Bayesian optimization over reduced space para 3 line 1-6, We implement the VAE model in PyTorch [33], an open source deep learning platform. The input to the VAE are images of size (batch, 1, 28, 28) while the output is of the same size. The input image database is constructed following a stochastic approach adopted in [34,35]. The method thresholds a Gaussian random field (GRF) to generate realistic binary representations of RVEs. The width for the hidden layer is set to be 10, namely, Z = R10. The encoder part consists of two convolutional and two dense layers [ there is an encoder trained to calculate a latent vector from a characteristic vector indicating a target composite characteristic]),
when a production condition vector indicating a production condition for obtaining the composite characteristic is received, by an inference unit, calculating, through a first regressor, a simulated latent vector simulating the latent vector from the input production condition vector (Xue page 3, Figure 2 Step 3,
PNG
media_image3.png
300
494
media_image3.png
Greyscale
[when a production condition vector indicating a production condition for obtaining the composite characteristic is received].
page 4 3.1. Bayesian optimization over reduced space para. 6,
The computational cost to evaluate the objective function in Eq. (5) is not cheap, motivating the use of BayesOpt, a class of machine-learning-based optimization methods. BayesOpt builds a probabilistic surrogate model for the objective function and queries the next data point by minimizing an expected loss function. The optimization loop is completely automatic and avoids subjective human decisions in trial-and-error design. In this work, we use noise free Gaussian process regression as the surrogate model for BayesOpt. We apply expected improvement, a classic acquisition function to determine the next data point to sample. The searching process in principle tries to gain a balance between exploitation and exploration. We point to [37,38] for comprehensive descriptions of BayesOp [calculating, through a first regressor,].
To further demonstrate the feasibility of optimizing over the latent space Z, we draw 200 samples from p(z), obtain decoded RVE images, and perform computational homogenization [a simulated latent vector simulating the latent vector from the input production condition vector].);
Burns and Xue are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Burns to incorporate the teachings of Xue to include using an encoder. Doing to learn the latent reduced representation to perform sequential optimization for multi-material design problems (Xue Abstract, Mechanical metamaterials are artificial structures that exhibit unusual mechanical properties at the macroscopic level due to architected geometric design at the microscopic level. With rapid advancement of multi-material 3D printing techniques, it is possible to design mechanical metamaterials by varying spatial distributions of different base materials within a representative volume element (RVE), which is then periodically arranged into a lattice structure. The design problem is challenging, however, considering the wide design space of potentially infinitely many configurations of multi-material RVEs. We propose an optimization framework that automates the design flow. We adopt variational autoencoder (VAE), a machine learning generative model to learn a latent, reduced representation of a given RVE configuration. The reduced design space allows to perform Bayesian optimization (BayesOpt), a sequential optimization strategy, for the multi-material design problems. In this work, we select two base materials with distinct elastic moduli and use the proposed optimization scheme to design a composite solid that achieves a prescribed set of macroscopic elastic moduli.).
and by the inference unit, calculating, through a decoder, a simulated characteristic vector simulating the characteristic vector from the simulated latent vector and an experimental condition vector indicating an experimental condition for measuring the composite characteristic (Kawamata page 2 Fig. 2,
PNG
media_image4.png
238
473
media_image4.png
Greyscale
[and by the inference unit, calculating, through a decoder,]]
Page 4, B. Visualization approach with CVAE
Visualization was carried out by inputting arbitrary magnetic material distribution X, condition Ytrue corresponding to X and intended condition Yopt., i.e., an intended magnetic energy level, into the learned CVAE. Ytrue and Yopt correspond to y and y´ in Fig. 3. Here, we select the shape X with low magnetic energy level among the dataset and input it to Encoder of CVAE with corresponding condition Ytrue. The latent variable z was sampled based on the mean and variance vector of Encoder outputs [a simulated characteristic vector simulating the characteristic vector from the simulated latent vector]. Next, we input the latent variable z with Yopt. Into Decoder. The value of each element of the decoder output X′ is a continuous value between 0 and 1 [and an experimental condition vector indicating an experimental condition for measuring the composite characteristic]).
Bruns and Kawamata are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Bruns to incorporate the teachings of Kawamata to include information an experimental conditions. Doing to enabling getting useful information form preferable design variables in accordance to the intended objective function making an effective design tools (C. Results of experiment para. 1, The above results of the numerical experiment indicate that CVAE integrate two kinds of information, i.e., the characteristics of the magnetic circuit shape and the magnetic energy value corresponding to the shape. In case the CVAE properly carries out training with the data prepared in advanced, it is shown that the proposed approach can visualize the direction of the modification of the magnetic circuit shape to satisfy the intended objective function values and constraints.
Page 4 IV. Conclusions The application of Conditional Variational Auto-Encoder to the design of fundamental linear magnetic circuit was successfully demonstrated. The developed method enables us to get useful information for obtaining preferable design variables in accordance with the intended objective function values by the designer, which can be an effective design tool. ).
Further claims 1-5 are contingent and are rejected as explained below as explained in MPEP 2111.04 II “The broadest reasonable interpretation of a method (or process) claim having contingent limitations requires only those steps that must be performed and does not include steps that are not required to be performed because the condition(s) precedent are not met.”. Claim 1 is a contingent method claim that has to contingent statements:
“in case that, upon producing a composite using multiple materials, there is an encoder trained to calculate a latent vector from a characteristic vector indicating a target composite characteristic,”
“when a production condition vector indicating a production condition for obtaining the composite characteristic is received, by an inference unit, calculating, through a first regressor, a simulated latent vector simulating the latent vector from the input production condition vector;”
The contingent statements above only requires that those steps that must be performed and does not include steps that are not required to be performed. Thus if the claim invention can be performed without the first or second condition happening as given the first limitation “in case that, …” the limitation requires that a composite material be produced there is an encoder. However there is no requirement for there to be encoder to produce a latent vector in the case that the composite is not required. Further the “when a production condition vector indicating a production condition for obtaining the composite characteristic is received” only occurs when a production condition is received. As explain in MPEP 2111.04 II no art is required to show any of the conditional limitations given the broadest reasonable interpretation of the limitation as there is condition where the method does not have to produce by an encoder a latent vector. Thus claim 1 can be rejected without the prior art and all dependent claims can therefore be rejected without the prior art.
Claim(s) 2-5 and 7-10 are rejected under 35 U.S.C. 103 as being unpatentable over Burns in view of Xue Kawamata and further in view of K. Mehta, S. S. Raju, M. Xiao, B. Wang, Y. Zhang and H. Y. Wong, "Improvement of TCAD Augmented Machine Learning Using Autoencoder for Semiconductor Variation Identification and Inverse Design," in IEEE Access, vol. 8, pp. 143519-143529, 2020, (“Mehta”).
Regarding claim 2 and analogous claim 7, Burns in view of Xue and Kawamata teach the method as recited in claim 1.
Burns, Xue and Kawamata are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Xue teaches further comprising:
when a characteristic vector is received, by the inference unit, calculating a latent vector from the characteristic vector through the encoder (Xue Page 3, Figure 1,
PNG
media_image2.png
675
1214
media_image2.png
Greyscale
3.1. Bayesian optimization over reduced space para 3 line 1-6, We implement the VAE model in PyTorch [33], an open source deep learning platform. The input to the VAE are images of size (batch, 1, 28, 28) while the output is of the same size. The input image database is constructed following a stochastic approach adopted in [34,35]. The method thresholds a Gaussian random field (GRF) to generate realistic binary representations of RVEs. The width for the hidden layer is set to be 10, namely, Z = R10. The encoder part consists of two convolutional and two dense layers [when a characteristic vector is received, by the inference unit, calculating a latent vector from the characteristic vector through the encoder; ]);
Burns does not explicitly teach and by the inference unit, calculating, through a second regressor, a simulated production condition vector simulating the production condition vector from the latent vector.
However Mehta teaches and by the inference unit, calculating, through a second regressor, a simulated production condition vector simulating the production condition vector from the latent vector (Metha page 143523 C. AUTOENCODER FOR VARIATION IDENTIFICATION para. 6, Fig. 8 represents the overall architecture of the machine learning model developed, that transforms linear and log scaled IV curve using autoencoder and predicts thickness using hidden layer neuron of the autoencoder. The scaler transformation of the IV curve to the autoencoder input in …,
PNG
media_image5.png
272
626
media_image5.png
Greyscale
[and by the inference unit, calculating, through a second regressor,]
Figure 8,
PNG
media_image6.png
612
1194
media_image6.png
Greyscale
[a simulated production condition vector simulating the production condition vector from the latent vector]).
Burns in view of Xue and are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Burns in view of Xue to incorporate the teachings of Mehta to include a linear regression to determine production condition. Doing to avoid overfitting, improve accuracy and generate sufficient data for ML model development (Mehta Abstract line 1-7, A machine learning (ML) model by combing two autoencoders and one linear regression model is proposed to avoid overfitting and to improve the accuracy of Technology Computer-Aided Design (TCAD)-augmented ML for semiconductor structural variation identification and inverse design, without using domain expertise. TCAD-augmented ML utilizes TCAD simulations to generate suficient data for ML model development when experimental data are inadequate. The ML model can then be used to identify semiconductor structural variation for given experimental electrical measurements. In this study, the variation of layer thicknesses in the p-i-n diode is used as a demonstration.).
Further claim 2 is a contingent method claim that depends on claim 1 as explained above, the limitation “when a characteristic vector is received, by the inference unit, calculating a latent vector from the characteristic vector through the encoder;” is contingent as similarly explained above the condition when the characteristic vector is not received does not require anything to happen. Therefore claim 2 can be rejected without the prior art (See MPEP 2111.04 II).
Regarding claim 3 and analogous claim 8, Burns in view of Xue and Kawamata and Mehta teach the method as recited in claim 2 and analogous 7.
Burns, Xue and Kawamata are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Burns and Mehta are combine in the same rational as set forth above with respect to claim 2 and analogous claim 7.
Further Kawamata taches further comprising:
before calculating the simulated latent vector, when a learning unit receives a characteristic vector for learning into the encoder, by the encoder, calculating a latent vector for learning from the characteristic vector for learning (Kawamata page 1, I. Introduction para. 3 line 12-19, CVAE is trained using two kinds of circuit characteristics, i.e., the shape (image) of magnetic material and the magnetic energy value of the targeted area as an objective function. As a result, inputting a shape of magnetic material corresponding to an arbitrary objective function value to the trained CVAE, we can derive an implicit suggestion of a shape change for increasing the objective function value, i.e., the useful information.
Page 1, B. Training process ,We use CVAE to extract features of targeted devices and visualize their information. Fig. 2. shows the conceptual diagram of training of CVAE. The CVAE used in the paper as training term is a network structure that converts an input image and conditions into latent variables of low-dimensional vectors and reproduces the variables and conditions to be identical to the input image. The part that converts the input image into latent variables is called an Encoder, and the part that restores it is called a Decoder. CVAE is a network that can impose
III. CONDITIONAL VARIATIONAL AUTO-ENCODER A. Formulation, We propose a visualization method of information using
CVAE. Based on the features learned from the data set, CVAE can generate data that meets additional conditions imposed during decompression process.
Fig. 1,
PNG
media_image7.png
261
470
media_image7.png
Greyscale
[before calculating the simulated latent vector, when a learning unit receives a characteristic vector for learning into the encoder,]
Fig. 2,
PNG
media_image8.png
231
466
media_image8.png
Greyscale
[by the encoder, calculating a latent vector for learning from the characteristic vector for learning;]);
when the learning unit receives a latent vector for learning and an experimental condition vector for learning into the decoder, by the decoder, calculating a simulated characteristic vector for learning from the latent vector for learning and the experimental condition vector for learning;
by the learning unit, calculating a loss representing a difference between the simulated characteristic vector for learning and the characteristic vector for learning;
and by the learning unit, performing optimization to update parameters of the encoder and the decoder to minimize the loss ((Kawamata Fig. 2,
PNG
media_image9.png
235
462
media_image9.png
Greyscale
[when the learning unit receives a latent vector for learning and an experimental condition vector for learning into the decoder, by the decoder, calculating a simulated characteristic vector for learning from the latent vector for learning and the experimental condition vector for learning; ]
page 2 B. Network structure
The network structure of CVAE used in this paper is shown in Fig. 4. The Encoder consists of two alternating layers of convolution and pooling followed by two full connected (FC) layers. The Decoder consists of two full connected layers followed by alternately three layers of convolution and two layers of upsampling. It is suggested that these layers are effective when deep generation model is used for image generation [9]. Parameters of the developed CVAE are summarized in Table 1.
Here, we assume q(z|X) to be multivariate normal distribution and p(z|X) to be the Bernoulli distribution. We conduct numerical experiments by using the developed CVAE with 30-dimensional latent space. The loss function is the binary cross entropy error between the training data and the output of the decoder [by the learning unit, calculating a loss representing a difference between the simulated characteristic vector for learning and the characteristic vector for learning]. The gradient descent algorithm adopts Adam [10], and the update equations are shown in (4).
PNG
media_image10.png
346
493
media_image10.png
Greyscale
[and by the learning unit, performing optimization to update parameters of the encoder and the decoder to minimize the loss]).
Further claim 3 is a contingent method claim that depends on claim 2 as explained above, the limitations:
“before calculating the simulated latent vector, when a learning unit receives a characteristic vector for learning into the encoder, by the encoder, calculating a latent vector for learning from the characteristic vector for learning;”
“when the learning unit receives a latent vector for learning and an experimental condition vector for learning into the decoder, by the decoder, calculating a simulated characteristic vector for learning from the latent vector for learning and the experimental condition vector for learning;”
Are contingent as similarly explained above the condition when the characteristic vector is not received or the latent vector is not received does not require anything to happen. Therefore claim 3 can be rejected without prior art (See MPEP 2111.04 II).
Regarding claim 4 and analogous claim 9, Burns in view of Xue and Kawamata and Mehta teach the method as recited in claim 2 and analogous 7.
Burns, Xue and Kawamata are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Burns and Mehta are combine in the same rational as set forth above with respect to claim 2 and analogous claim 7.
Kawamata teaches further comprising: before calculating the simulated latent vector, by a learning unit, preparing a plurality of learning data, each of which includes a latent vector for learning derived from a characteristic vector for learning by the encoder and a production condition vector for learning corresponding to the characteristic vector for learning (Kawamata page 1, B. Training process, We use CVAE to extract features of targeted devices and visualize their information. Fig. 2. shows the conceptual diagram of training of CVAE. The CVAE used in the paper as training term is a network structure that converts an input image and conditions into latent variables of low-dimensional vectors and reproduces the variables and conditions to be identical to the input image. The part that converts the input image into latent variables is called an Encoder, and the part that restores it is called a Decoder. CVAE is a network that can impose conditions during both learning and inference, and conduct conditional learning.
Page 2 III. CONDITIONAL VARIATIONAL AUTO-ENCODER A. Formulation, We propose a visualization method of information using
CVAE. Based on the features learned from the data set, CVAE can generate data that meets additional conditions imposed during decompression process.
Fig. 1,
PNG
media_image7.png
261
470
media_image7.png
Greyscale
[each of which includes a latent vector for learning derived from a characteristic vector for learning by the encoder and a production condition vector for learning corresponding to the characteristic vector for learning]
Fig. 2,
PNG
media_image8.png
231
466
media_image8.png
Greyscale
[before calculating the simulated latent vector, by a learning unit, preparing a plurality of learning data,]);
Further Mehta teaches by the learning unit, preparing a prototype of a second regressor in which the latent vector for learning is used as an input, the production condition vector for learning is used as an output, and a weight for the latent vector for learning is not determined;
and completing the second regressor by deriving a weight for the latent vector for learning using the learning data (Mehta page 143521 B. Linear regression and overfitting para 2 line1-10, Firstly, training data in Set 1 is used to train an ML model (dubbed as Model-R-1) using linear regression in the Scikit-learn library [14] . Here tn, ti, and tp are Oy and Xtrain_scaled are Xinput . This model is then validated by the test data in Set 1 [by the learning unit, preparing a prototype of a second regressor in which the latent vector for learning is used as an input,]. Fig. 3 shows that the model has an excellent prediction on the test data tn, ti, and tp. However, when the model is used to predict the thicknesses of the devices in Set 2, the performance is very bad (Fig. 4). It may wrongly predict thickness as large as 106nm or even predict negative thicknesses (e.g. tn C tp < -50000nm).
Page 143522 para 3, It is also useful to investigate how the model will improv if it is allowed to “see” Set 2 (i.e. to be aware of the existence of doping concentration variation) during training. Set 1 and Set 2 are combined and 90% of the data are used to train a new model (dubbed as Model-R-1-2). 10% of the combined dataset are used for validation. Note that the outputs are still only tn, ti; and tp. Doping concentration is not used in the training. Therefore, the new model still does not have information on how doping concentration will modify the IV curves. Fig. 5 shows the validation result of Model-R-1-2 by using the testing data of the combined dataset. The model now gives much more reasonable predictions on tn, ti; and tp for Set 2 curves, which are supposed to be 200 nm, 10 nm, and 200 nm respectively [the production condition vector for learning is used as an output,] (i.e. before training the model is prepared and no weights are determined for the latent vector).
Page 143523 III. Machine Learning Models C. Autoencoder for variation Identification para 7,,
PNG
media_image11.png
486
523
media_image11.png
Greyscale
[and completing the second regressor by deriving a weight for the latent vector for learning using the learning data.]).
Regarding claim 5 and analogous claim 10, Burns in view of Xue and Kawamata and Mehta teach the method as recited in claim 2 and analogous 7.
Burns, Xue and Kawamata are combine in the same rational as set forth above with respect to claim 1 and analogous claim 11.
Burns and Mehta are combine in the same rational as set forth above with respect to claim 2 and analogous claim 7.
Xue teaches before calculating the simulated latent vector, by a learning unit, preparing a plurality of learning data, each of which includes a production condition vector for learning and a latent vector for learning derived from a characteristic vector for learning corresponding to the production condition vector for learning by the encoder ((Xue
Page 2, 2. Computational homogenization for linear elasticity para. 4, As shown in Fig. 1, we summarize the computational homogenization problem as to find out the macroscopic elastic moduli {E, ν, G} given an image-like description (binary pixel values indicating base material information) of a RVE. The computational homogenization technique requires the use of the Finite Element Method (FEM). We choose the open source FEM software FEniCS [31] for the implementation.
Page 3, Figure 1,
PNG
media_image2.png
675
1214
media_image2.png
Greyscale
[each of which includes a production condition vector for learning and a latent vector for learning derived from a characteristic vector for learning corresponding to the production condition vector for learning by the encoder;]
3.1. Bayesian optimization over reduced space para 3, We implement the VAE model in PyTorch [33], an open source deep learning platform. The input to the VAE are images of size (batch, 1, 28, 28) while the output is of the same size. The input image database is constructed following a stochastic approach adopted in [34,35]. The method thresholds a Gaussian random field (GRF) to generate realistic binary representations of RVEs. The width for the hidden layer is set to be 10, namely, Z = R10. The encoder part consists of two convolutional and two dense layers Similarly, the decoder consists of two dense layers followed by two deconvolutional layers. Right before the output layer we add an additional constraint to force diagonal symmetry on the output tensors. To generate new samples, we only need the decoder model. We draw random variables from a unit Gaussian distribution as the input to the decoder and threshold the decoded images to reconstruct the binary composition. Fig. 2 shows a gallery of generated RVE samples. A more detailed description on VAEs can be found in Appendix B. [before calculating the simulated latent vector, by a learning unit, preparing a plurality of learning data,]));
by the learning unit, preparing a prototype of a first regressor in which the production condition vector for learning is used an input, the latent vector for learning is used as an output, and a weight of the production condition vector for learning is not determined; and completing the first regressor by deriving a weight of the production condition vector for learning using the learning data (Xue page 4, 3.1. Bayesian optimization over reduced space Para. 5, The computational cost to evaluate the objective function in Eq. (5) is not cheap, motivating the use of BayesOpt, a class of machine-learning-based optimization methods [by the learning unit, preparing a prototype of a first regressor in which the production condition vector for learning is used an input,]. BayesOpt builds a probabilistic surrogate model for the objective function and queries the next data point by minimizing an expected loss function. The optimization loop is completely automatic and avoids subjective human decisions in trial-and-error design. In this work, we use noise free Gaussian process regression as the surrogate model for BayesOpt. We apply expected improvement, a classic acquisition function to determine the next data point to sample. The searching process in principle tries to gain a balance between exploitation and exploration. We point to [37,38] for comprehensive descriptions of BayesOpt [, the latent vector for learning is used as an output, and a weight of the production condition vector for learning is not determined]
page 5 3.2 Experimental Validation para. 3,
PNG
media_image12.png
263
520
media_image12.png
Greyscale
where wE and wν are two constants reflecting a bias on the optimization targets [by deriving a weight of the production condition vector for learning using the learning data]).
Pertinent Prior Art
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Amer et al. (US20190094124A1) – teaches identifying analyzing corrosion based on different conditions by using a neural network.
Strope et al. (US11072133B1) – teaches manufacturing composite structures and using machine learning to determine adjustments in the production.
Dong, Yuan, et al. "Inverse Structural Design of Graphene/Boron Nitride Hybrids by Regressional GAN." arXiv preprint arXiv:1908.07959 – teaches a regressional adversarial network that enables autonomously generate graphene boron-nitride hybrid with any given bandgaps.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ALFREDO CAMPOS whose telephone number is (571)272-4504. The examiner can normally be reached 7:00 - 4:00 pm M - F.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Michael J. Huntley can be reached at (303) 297-4307. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/ALFREDO CAMPOS/Examiner, Art Unit 2129
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129
Prosecution Insights