DETAILED ACTION
This action is in response to claims filed 06 October 2023 for application 18377520. Currently claims 1-20 are pending.
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 .
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(d):
(d) REFERENCE IN DEPENDENT FORMS.—Subject to subsection (e), a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers.
The following is a quotation of pre-AIA 35 U.S.C. 112, fourth paragraph:
Subject to the following paragraph [i.e., the fifth paragraph of pre-AIA 35 U.S.C. 112], a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers.
Claims 18 and 20 are rejected under 35 U.S.C. 112(d) or pre-AIA 35 U.S.C. 112, 4th paragraph, as being of improper dependent form for failing to further limit the subject matter of the claim upon which it depends, or for failing to include all the limitations of the claim upon which it depends. Claims 18 and 20 depend upon themselves respectively. They will be interpreted as depending on claim 17 and 19 respectively. Applicant may cancel the claim(s), amend the claim(s) to place the claim(s) in proper dependent form, rewrite the claim(s) in independent form, or present a sufficient showing that the dependent claim(s) complies with the statutory requirements.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claim(s) 1-6, 9, 11-13, and 15-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Yang et al. (Taking materials dynamics to new extremes using machine learning interatomic potentials) in view of Liu et al. (Deep Neural Network Ensembles against Deception: Ensemble Diversity, Accuracy and Robustness).
Regarding claims 1, 17 and 19, Yang discloses:
An iterative machine learning interatomic potential (MLIP) training method comprising:
training a first multiplicity of first MLIP models in a first iteration of a training loop (“Typical strategies for efficiently sampling in active learning. (A) Based on algorithms e.g., Gaussian regression, density functional theory (DFT) calculations are usually applied on configurations with high uncertainty. (B) Uncertainty can also be evaluated by several parallel models; the standard derivation (STD) of N predictions reflects the uncertainty of selected configurations. (C) Global optimizations are also applied to select important structures for DFT calculations[52].” Fig 2, “In this scenario, a raw MLIP was first constructed by a small dataset, then MD simulations were run using this potential to generate more configurations. The configurations with high uncertainty are selected and then re-labeled by DFT calculations and updated the training dataset [Figure 2A]. Although it often takes several iterations to achieve the acceptable learning dataset, active learning can effectively reduce the size of the training dataset without loss of training accuracy” p4 ¶1);
a second multiplicity of second MLIP models p4 ¶1
iteratively trained MLIP configured to predict one or more values of a material (“The aim of an MLIP is to map the configuration of a system in real space into its potential energy surface (PES). Based on a set of discrete points on PES (generated by DFT calculations), ML regression algorithms are applied to learn a smooth PES. We can then predict the potential energy E of a given configuration. The total force Fi acting on individual atom i in this system can be expressed as Fi = -, where ri is the position of atom i. MD simulations can be performed similarly to those with classical interatomic potentials. The general ingredients of MLIPs include the representation for local atomic environments, learning databases, regression models, and evaluation algorithms [Figure 1]. We will review these key aspects in this section.” P3 ¶2).
However, Yang does not explicitly disclose: training a second multiplicity of second … models in a second iteration of the training loop in parallel with the first training step; and
combining the first … models and the second … models.
Liu teaches: training a second multiplicity of second … models in a second iteration of the training loop in parallel with the first training step; and combining the first … models and the second … models (“Ensemble Diversity, Accuracy and Robustness. The ensemble DNN prediction is performed by combining the individual predictions from all members of the committee via a consensus method. The architecture of an ensemble learner can be parallel, parallel hierarchical (e.g., Boosting), parallel cascading, gated parallel, to name a few. The consensus methods of an ensemble learner can be as simple as majority voting, sum, mean, median, or more complex, such as weighted averaging, voting by rank using a set of statistic metrics in terms of prediction confidence, rank score, rank confidence or more abstract ones. The best scenario is when all members of an ensemble committee of size N can learn and predict with uncorrelated errors.” P1 ¶2).
Yang and Liu are in the same field of endeavor of machine learning models and are analogous. Yang discloses MLIP models using various ML models trained in an iterative manner. Liu teaches ensembles of models trained in parallel and selected using various methods. It would have been obvious to one of ordinary skill in the art before the effective filing date to modify the iterative MLIP model of Yang with the known ensemble and parallel model structure as taught by Liu to yield predictable results of more accurate models that are more robust to errors.
Regarding claim 2, Yang discloses: The iterative MLIP training method of claim 1 further comprising predicting the one or more values of the material with the iteratively trained MLIP (“The aim of an MLIP is to map the configuration of a system in real space into its potential energy surface (PES). Based on a set of discrete points on PES (generated by DFT calculations), ML regression algorithms are applied to learn a smooth PES. We can then predict the potential energy E of a given configuration. The total force Fi acting on individual atom i in this system can be expressed as Fi = -, where ri is the position of atom i. MD simulations can be performed similarly to those with classical interatomic potentials. The general ingredients of MLIPs include the representation for local atomic environments, learning databases, regression models, and evaluation algorithms [Figure 1]. We will review these key aspects in this section.” P3 ¶2, Fig 2).
Regarding claim 3, Yang discloses: The iterative MLIP training method of claim 1, wherein the one or more values include total energy, atomic forces, atomic stresses, atomic charges, and/or polarization (“The aim of an MLIP is to map the configuration of a system in real space into its potential energy surface (PES). Based on a set of discrete points on PES (generated by DFT calculations), ML regression algorithms are applied to learn a smooth PES. We can then predict the potential energy E of a given configuration. The total force Fi acting on individual atom i in this system can be expressed as Fi = -, where ri is the position of atom i. MD simulations can be performed similarly to those with classical interatomic potentials. The general ingredients of MLIPs include the representation for local atomic environments, learning databases, regression models, and evaluation algorithms [Figure 1]. We will review these key aspects in this section.” P3 ¶2).
Regarding claim 4, Yang does not explicitly disclose, however Liu teaches: The iterative MLIP training method of claim 1 further comprising calling a third MLIP model when a predicted confidence of the first and/or second MLIP models falls below a predicted confidence threshold (ensemble selection p7 §IV).
Regarding claim 5, Yang does not explicitly disclose, however Liu teaches: The iterative MLIP training method of claim 1 further comprising calling a third MLIP model when a predicted uncertainty of the first and/or second MLIP models exceeds a predicted uncertainty threshold (ensemble selection p7 §IV).
Regarding claim 6, Yang does not explicitly disclose, however Liu teaches: The iterative MLIP training method of claim 1 wherein the first MLIP models use a first set of hyperparameters and the second MLIP models use a second set of hyperparameters different than the first set of hyperparameters (“First, we define the concept of ensemble diversity by examining three types of diversity used in constructing classification ensembles: (i) the model diversity by their difference in DNN algorithm, neural network structure/topology, and hyperparameters used for classifier training and prediction; (ii) the model diversity by their disagreement on negative examples, aiming to promote failure independence of ensemble member models, to increase the overall performance (accuracy) of ensemble prediction, and the robustness of ensemble against deception; and (iii) the model diversity during training by altering the way that each individual learner traverses the hypothesis space, leading different classifiers to converge to different hypotheses within the classification manifold. ”, p2 ¶3).
Regarding claim 9, Yang does not explicitly disclose, however Liu teaches: The iterative MLIP training method of claim 1 further comprising selecting the iterative MLIP model from the model of the first and second MLIP models having the lowest error of the errors of the first and second MLIP models (ensemble selection p7 §IV).
Regarding claim 11, Yang discloses: The interactive MLIP training method of claim 1, wherein the combining step accounts for the atomic changes in energies (“Representation of atomic structures entails quantifying local structural information in certain mathematical expressions, named descriptors or fingerprints[53]. In order to simulate large systems, the total energy is expressed as a linear combination of the sum of local energy contributions from all the atoms. Similarly, the fingerprints can also be simplified as a linear summation of local atomic environments, i.e., D = ΣDi, where D is the fingerprints of the system, and Di is the fingerprint of atom i. This assumption greatly improves the efficiency of MLIPs[54,55].” P4 ¶2).
Regarding claim 12, Yang discloses: The iterative MLIP training method of claim 1, wherein the first training step produces a first machine learning system and the second training step produces a second machine learning system, and further comprising determining one or more instabilities in response to the first and second machine learning systems (“Typical strategies for efficiently sampling in active learning. (A) Based on algorithms e.g., Gaussian regression, density functional theory (DFT) calculations are usually applied on configurations with high uncertainty. (B) Uncertainty can also be evaluated by several parallel models; the standard derivation (STD) of N predictions reflects the uncertainty of selected configurations. (C) Global optimizations are also applied to select important structures for DFT calculations[52].” Fig 2).
Regarding claim 13, Yang discloses: The iterative MLIP training method of claim 12 further comprising learning from the one or more instabilities when training a … MLIP model (“Typical strategies for efficiently sampling in active learning. (A) Based on algorithms e.g., Gaussian regression, density functional theory (DFT) calculations are usually applied on configurations with high uncertainty. (B) Uncertainty can also be evaluated by several parallel models; the standard derivation (STD) of N predictions reflects the uncertainty of selected configurations. (C) Global optimizations are also applied to select important structures for DFT calculations[52].”).
Liu discloses: a third model (“Ensemble Diversity, Accuracy and Robustness. The ensemble DNN prediction is performed by combining the individual predictions from all members of the committee via a consensus method. The architecture of an ensemble learner can be parallel, parallel hierarchical (e.g., Boosting), parallel cascading, gated parallel, to name a few. The consensus methods of an ensemble learner can be as simple as majority voting, sum, mean, median, or more complex, such as weighted averaging, voting by rank using a set of statistic metrics in terms of prediction confidence, rank score, rank confidence or more abstract ones. The best scenario is when all members of an ensemble committee of size N can learn and predict with uncorrelated errors.” P1 ¶2)
Regarding claim 15, Yang discloses: The iterative MLIP training method of claim 1, wherein the first training step produces a first machine learning system, and further comprising generating at least one starting structure for an active learning scheme in the second training step in response to the first machine learning system (“Typical strategies for efficiently sampling in active learning. (A) Based on algorithms e.g., Gaussian regression, density functional theory (DFT) calculations are usually applied on configurations with high uncertainty. (B) Uncertainty can also be evaluated by several parallel models; the standard derivation (STD) of N predictions reflects the uncertainty of selected configurations. (C) Global optimizations are also applied to select important structures for DFT calculations[52].” Fig 2).
Regarding claim 16, Yang discloses: The iterative MLIP training method of claim 1 further comprising removing datapoints based on a thermodynamic relevance (“The configurations with high uncertainty are selected and then re-labeled by DFT calculations and updated the training dataset [Figure 2A]. Although it often takes several iterations to achieve the acceptable learning dataset, active learning can effectively reduce the size of the training dataset without loss of training accuracy. The most important process in active learning is re-sampling. However, most ML regression algorithms cannot provide the uncertainty directly. To solve this problem, random sampling methods (e.g., Bootstrapping) have been used to evaluate the uncertainty [Figure 2B]. Even so, not all important metastable structures can be sampled. Therefore, we often enrich the learning dataset with the help of some global optimization method[52], which can be more efficient in sampling metastable structures along the phase transition pathways [Figure 2C].” p4 ¶1).
Regarding claim 18, Yang discloses: The iterative MLIP training method of claim 18, wherein the first and second MLIP models are first and second Gaussian Process (GP) based MLIPs (“Typical strategies for efficiently sampling in active learning. (A) Based on algorithms e.g., Gaussian regression, density functional theory (DFT) calculations are usually applied on configurations with high uncertainty. (B) Uncertainty can also be evaluated by several parallel models; the standard derivation (STD) of N predictions reflects the uncertainty of selected configurations. (C) Global optimizations are also applied to select important structures for DFT calculations[52].” Fig 2, “In this scenario, a raw MLIP was first constructed by a small dataset, then MD simulations were run using this potential to generate more configurations. The configurations with high uncertainty are selected and then re-labeled by DFT calculations and updated the training dataset [Figure 2A]. Although it often takes several iterations to achieve the acceptable learning dataset, active learning can effectively reduce the size of the training dataset without loss of training accuracy” p4 ¶1).
Regarding claim 20, Yang discloses: The iterative MLIP training method of claim 20 , wherein the first and second MLIP models are first and second deep learning based MLIP models (Fig 4).
Claim(s) 7-8 and 10 are rejected under 35 U.S.C. 103 as being unpatentable over Yang in view of Liu and further in view of Kornbluth et al. (US 20220100931).
Regarding claim 7, Yang does not explicitly disclose, however, Kornbluth teaches: The iterative MLIP training method of claim 1, wherein the first MLIP models use a first starting atomic structure and the second MLIP models use a second starting atomic structure different than the first starting atomic structure (“Even for classical potentials, scaling MD simulations across multiple CPUs is a well-investigated problem. One such solution is the Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS). LAMMPS operates by distributing atoms across multiple CPUs by assigning each CPU the ownership of any atoms within a volume in physical space (e.g., a 2×2×3 nm cube). Instead of retaining information about every atom in the system, the CPU only operates on information regarding the atoms within its corresponding volume plus those atoms adjacent to the volume. LAMMPS utilizes various algorithms to identify the atomic neighbors (referred to as neighbor lists) for each CPU. These atomic neighbors include ghost atoms, which refers to atoms that are owned by a different CPU but are close enough to require the CPU itself to have information about its position. With this information, the forces can be predicted individually on each CPU, with minimal communication between CPUs only when atoms move from the ownership (i.e., the computational domain) of one CPU to that of another CPU.” [0033]).
Yang, Liu and Kornbluth are in the same field of endeavor of machine learning models and are analogous. Yang discloses MLIP models using various ML models trained in an iterative manner. Liu teaches ensembles of models trained in parallel and selected using various methods. Kornbluth discloses an MLIP model distributed over multiple processors having separate atoms with different parameters on each processor. It would have been obvious to one of ordinary skill in the art before the effective filing date to modify the ensemble iterative MLIP model of Yang and Liu with the known MLIP model having different atoms and atomic environments spread across processors and models as taught by Kornbluth to yield predictable results of reducing communication requirements.
Regarding claim 8, Yang does not explicitly disclose, however, Kornbluth teaches: The iterative MLIP training method of claim 1, wherein the first MLIP models use a first chemical composition and the second MLIP models use a second chemical composition different than the first chemical composition (“Even for classical potentials, scaling MD simulations across multiple CPUs is a well-investigated problem. One such solution is the Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS). LAMMPS operates by distributing atoms across multiple CPUs by assigning each CPU the ownership of any atoms within a volume in physical space (e.g., a 2×2×3 nm cube). Instead of retaining information about every atom in the system, the CPU only operates on information regarding the atoms within its corresponding volume plus those atoms adjacent to the volume. LAMMPS utilizes various algorithms to identify the atomic neighbors (referred to as neighbor lists) for each CPU. These atomic neighbors include ghost atoms, which refers to atoms that are owned by a different CPU but are close enough to require the CPU itself to have information about its position. With this information, the forces can be predicted individually on each CPU, with minimal communication between CPUs only when atoms move from the ownership (i.e., the computational domain) of one CPU to that of another CPU.” [0033]).
Regarding claim 10, Yang does not explicitly disclose, however, Kornbluth teaches: The iterative MLIP training method of claim 1, wherein the combining step accounts for an atomic environment overlap (“Even for classical potentials, scaling MD simulations across multiple CPUs is a well-investigated problem. One such solution is the Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS). LAMMPS operates by distributing atoms across multiple CPUs by assigning each CPU the ownership of any atoms within a volume in physical space (e.g., a 2×2×3 nm cube). Instead of retaining information about every atom in the system, the CPU only operates on information regarding the atoms within its corresponding volume plus those atoms adjacent to the volume. LAMMPS utilizes various algorithms to identify the atomic neighbors (referred to as neighbor lists) for each CPU. These atomic neighbors include ghost atoms, which refers to atoms that are owned by a different CPU but are close enough to require the CPU itself to have information about its position. With this information, the forces can be predicted individually on each CPU, with minimal communication between CPUs only when atoms move from the ownership (i.e., the computational domain) of one CPU to that of another CPU.” [0033]).
Claim 14 is rejected under 35 U.S.C. 103 as being unpatentable over Yang in view of Liu and further in view of Johnson et al. (US 20180110973).
Regarding claim 14, Yang does not explicitly disclose, however, Johnson teaches: The iterative MLIP training method of claim 1 further comprising terminating the training loop when one or more of the models of the first and second MLIP models is not at least near a Pareto front (“In embodiments, the generation limit can be any number of generations. In an example embodiment, 200 generations is the limit, as empirical data finds this to enable sufficient iterations for convergent runs to terminate while also enabling non-convergent runs to explore the solution space and improve Pareto front estimates.” [0106]).
Yang, Liu and Kornbluth are in the same field of endeavor of machine learning models and are analogous. Yang discloses MLIP models using various ML models trained in an iterative manner. Liu teaches ensembles of models trained in parallel and selected using various methods. Johnson teaches stopping conditions and failure to converge to a Pareto optimal solution. It would have been obvious to one of ordinary skill in the art before the effective filing date to modify the ensemble iterative MLIP model of Yang and Liu with the known stopping conditions as taught by Johnson to yield predictable results of the most information if no solution is found.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Vandenhaute et al. (Machine learning potentials for metal-organic frameworks using an incremental learning approach) and Chan et al. (Machine Learning Classical Interatomic Potentials for Molecular Dynamics from First-Principles Training Data) disclose MLIP models and usage.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ERIC NILSSON whose telephone number is (571)272-5246. The examiner can normally be reached M-F: 7-3.
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, James Trujillo can be reached at (571)-272-3677. 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.
/ERIC NILSSON/ Primary Examiner, Art Unit 2151
Prosecution Insights