Uncertainty-driven dynamics for active learning of interatomic potentials

Nature Computational Science, Journal Year: 2023, Volume and Issue: 3(3), P. 230 - 239

Published: March 6, 2023

Machine learning (ML) models, if trained to data sets of high-fidelity quantum simulations, produce accurate and efficient interatomic potentials. Active (AL) is a powerful tool iteratively generate diverse sets. In this approach, the ML model provides an uncertainty estimate along with its prediction for each new atomic configuration. If passes certain threshold, then configuration included in set. Here we develop strategy more rapidly discover configurations that meaningfully augment training The uncertainty-driven dynamics active (UDD-AL), modifies potential energy surface used molecular simulations favor regions space which there large uncertainty. performance UDD-AL demonstrated two AL tasks: sampling conformational glycine promotion proton transfer acetylacetone. method shown efficiently explore chemically relevant space, may be inaccessible using regular dynamical at target temperature conditions.

Language: Английский

---

Robust training of machine learning interatomic potentials with dimensionality reduction and stratified sampling

npj Computational Materials, Journal Year: 2024, Volume and Issue: 10(1)

Published: Feb. 26, 2024

Abstract Machine learning interatomic potentials (MLIPs) enable accurate simulations of materials at scales beyond that accessible by ab initio methods and play an increasingly important role in the study design materials. However, MLIPs are only as robust data on which they trained. Here, we present DImensionality-Reduced Encoded Clusters with sTratified (DIRECT) sampling approach to select a training set structures from large complex configuration space. By applying DIRECT Materials Project relaxation trajectories dataset over one million 89 elements, develop improved 3-body graph network (M3GNet) universal potential extrapolates more reliably unseen structures. We further show molecular dynamics (MD) M3GNet can be used instead expensive MD rapidly create space for target systems. combined this scheme reliable moment tensor titanium hydrides without need iterative augmentation This work paves way high-throughput development across any compositional complexity.

Language: Английский

---

Complexity of many-body interactions in transition metals via machine-learned force fields from the TM23 data set

npj Computational Materials, Journal Year: 2024, Volume and Issue: 10(1)

Published: May 7, 2024

Abstract This work examines challenges associated with the accuracy of machine-learned force fields (MLFFs) for bulk solid and liquid phases d -block elements. In exhaustive detail, we contrast performance force, energy, stress predictions across transition metals two leading MLFF models: a kernel-based atomic cluster expansion method implemented using sparse Gaussian processes (FLARE), an equivariant message-passing neural network (NequIP). Early present higher relative errors are more difficult to learn late platinum- coinage-group elements, this trend persists model architectures. Trends in complexity interatomic interactions different revealed via comparison representations many-body order angular resolution. Using arguments based on perturbation theory occupied unoccupied states near Fermi level, determine that large, sharp density both above below level early leads complex, harder-to-learn potential energy surface these metals. Increasing fictitious electronic temperature (smearing) modifies sensitivity forces makes metal easier learn. illustrates capturing intricate properties metallic bonding current MLFFs provides reference data set metals, aimed at benchmarking improving development emerging approximations.

Language: Английский

---

Accurate energy barriers for catalytic reaction pathways: an automatic training protocol for machine learning force fields

npj Computational Materials, Journal Year: 2023, Volume and Issue: 9(1)

Published: Oct. 4, 2023

Abstract We introduce a training protocol for developing machine learning force fields (MLFFs), capable of accurately determining energy barriers in catalytic reaction pathways. The is validated on the extensively explored hydrogenation carbon dioxide to methanol over indium oxide. With help active learning, final field obtains within 0.05 eV Density Functional Theory. Thanks computational speedup, not only do we reduce cost routine in-silico tasks, but also find an alternative path previously established rate-limiting step, with 40% reduction activation energy. Furthermore, illustrate importance finite temperature effects and compute free barriers. transferability demonstrated experimentally relevant, yet unexplored, top-layer reduced oxide surface. ability MLFFs enhance our understanding studied catalysts underscores need fast accurate alternatives direct ab-initio simulations.

Language: Английский

---

Data Generation for Machine Learning Interatomic Potentials and Beyond

Chemical Reviews, Journal Year: 2024, Volume and Issue: 124(24), P. 13681 - 13714

Published: Nov. 21, 2024

The field of data-driven chemistry is undergoing an evolution, driven by innovations in machine learning models for predicting molecular properties and behavior. Recent strides ML-based interatomic potentials have paved the way accurate modeling diverse chemical structural at atomic level. key determinant defining MLIP reliability remains quality training data. A paramount challenge lies constructing sets that capture specific domains vast space. This Review navigates intricate landscape essential components integrity data ensure extensibility transferability resulting models. We delve into details active learning, discussing its various facets implementations. outline different types uncertainty quantification applied to atomistic acquisition correlations between estimated true error. role samplers generating informative structures highlighted. Furthermore, we discuss via modified surrogate potential energy surfaces as innovative approach diversify also provides a list publicly available cover

Language: Английский

---

Efficiency, accuracy, and transferability of machine learning potentials: Application to dislocations and cracks in iron

Acta Materialia, Journal Year: 2024, Volume and Issue: 270, P. 119788 - 119788

Published: Feb. 29, 2024

Machine learning interatomic potentials (ML-IAPs) enable quantum-accurate, classical molecular dynamics simulations of large systems, beyond reach density functional theory (DFT). Yet, their efficiency and ability to predict systems larger than DFT supercells are not fully explored, posing a question regarding transferability large-scale with defects (e.g. dislocations, cracks). Here, we apply three-step validation approach body-centered-cubic iron. First, accuracy assessed by optimizing ML-IAPs based on four state-of-the-art ML packages. The Pareto front computational speed versus testing root-mean-square-error (RMSE) is computed. Second, benchmark properties relevant plasticity fracture evaluated. Their relative (Q) respect found correlate RMSE. Third, dislocations cracks investigated using per-atom model uncertainty quantification. core structures Peierls barriers screw, M111 three edge compared DFT. Traction-separation curve critical stress intensity factor (KIc) also predicted. Cleavage the pre-existing crack plane be zero-temperature atomistic mechanism pure iron under mode-I loading, independent package training database. Quantitative predictions dislocation glide paths KIc can sensitive database, package, cutoff radius, limited accuracy. Our results highlight importance validating indicators Moreover, significant speed-ups achieved most efficient ML-IAP yet assessment should performed care.

Language: Английский

---

Uncertainty-biased molecular dynamics for learning uniformly accurate interatomic potentials

npj Computational Materials, Journal Year: 2024, Volume and Issue: 10(1)

Published: April 29, 2024

Abstract Efficiently creating a concise but comprehensive data set for training machine-learned interatomic potentials (MLIPs) is an under-explored problem. Active learning, which uses biased or unbiased molecular dynamics (MD) to generate candidate pools, aims address this objective. Existing and MD-simulation methods, however, are prone miss either rare events extrapolative regions—areas of the configurational space where unreliable predictions made. This work demonstrates that MD, when by MLIP’s energy uncertainty, simultaneously captures regions events, crucial developing uniformly accurate MLIPs. Furthermore, exploiting automatic differentiation, we enhance bias-forces-driven MD with concept bias stress. We employ calibrated gradient-based uncertainties yield MLIPs similar or, sometimes, better accuracy than ensemble-based methods at lower computational cost. Finally, apply uncertainty-biased alanine dipeptide MIL-53(Al), generating represent both spaces more accurately models trained conventional MD.

Language: Английский

---

An overview about neural networks potentials in molecular dynamics simulation

International Journal of Quantum Chemistry, Journal Year: 2024, Volume and Issue: 124(11)

Published: May 21, 2024

Abstract Ab‐initio molecular dynamics (AIMD) is a key method for realistic simulation of complex atomistic systems and processes in nanoscale. In AIMD, finite‐temperature dynamical trajectories are generated by using forces computed from electronic structure calculations. with high numbers components typical AIMD run computationally demanding. On the other hand, machine learning (ML) subfield artificial intelligence that consist set algorithms show experience use input output data where capable analysing predicting future. At present, main application ML techniques atomic simulations development new interatomic potentials to correctly describe potential energy surfaces (PES). This technique constant progress since its inception around 30 years ago. The combine advantages classical methods, is, efficiency simple functional form accuracy first principles this article we review evolution four generations some their most notable applications. focuses on MLPs based neural networks. Also, present state art topic future trends. Finally, report results scientometric study (covering period 1995–2023) about impact applied simulations, distribution publications geographical regions hot topics investigated literature.

Language: Английский

---

Targeted Transferable Machine-Learned Potential for Linear Alkanes Trained on C₁₄H₃₀ and Tested for C₄H₁₀ to C₃₀H₆₂

Journal of Chemical Theory and Computation, Journal Year: 2025, Volume and Issue: unknown

Published: March 27, 2025

Given the great importance of linear alkanes in fundamental and applied research, an accurate machine-learned potential (MLP) would be a major advance computational modeling these hydrocarbons. Recently, we reported novel, many-body permutationally invariant model that was trained specifically for 44-atom hydrocarbon C14H30 on roughly 250,000 B3LYP energies (Qu, C.; Houston, P. L.; Allison, T.; Schneider, B. I.; Bowman, J. M. Chem. Theory Comput. 2024, 20, 9339–9353). Here, demonstrate accuracy transferability this ranging from butane C4H10 up to C30H62. Unlike other approaches aim universal applicability, present approach is targeted alkanes. The mean absolute error (MAE) energy ranges 0.26 kcal/mol rises 0.73 C30H62 over range 80 600 These values are unprecedented transferable potentials indicate high performance potential. conformational barriers shown excellent agreement with high-level ab initio calculations pentane, largest alkane which such have been reported. Vibrational power spectra molecular dynamics presented briefly discussed. Finally, evaluation time vary linearly number atoms.

Language: Английский

---

Generator of Neural Network Potential for Molecular Dynamics: Constructing Robust and Accurate Potentials with Active Learning for Nanosecond-Scale Simulations

Journal of Chemical Theory and Computation, Journal Year: 2025, Volume and Issue: unknown

Published: April 7, 2025

Neural network potentials (NNPs) enable large-scale molecular dynamics (MD) simulations of systems containing >10,000 atoms with the accuracy comparable to ab initio methods and play a crucial role in material studies. Although NNPs are valuable for short-duration MD simulations, maintaining stability long-duration remains challenging due uncharted regions potential energy surface (PES). Currently, there is no effective methodology address this issue. To overcome challenge, we developed an automatic generator robust accurate based on active learning (AL) framework. This provides fully integrated solution encompassing initial data set creation, NNP training, evaluation, sampling additional structures, screening, labeling. Crucially, our approach uses strategy that focuses generating unstable structures short interatomic distances, combined screening efficiently samples these configurations distances structural features. greatly enhances simulation stability, enabling nanosecond-scale simulations. We evaluated performance terms its physical properties by applying it liquid propylene glycol (PG) polyethylene (PEG). The generated stable 20 ns. predicted properties, such as density self-diffusion coefficient, show excellent agreement experimental values. work represents remarkable advance generation organic materials, paving way complex systems.

Language: Английский

---