DeePMD-kit v2: A software package for deep potential models

The Journal of Chemical Physics, Journal Year: 2023, Volume and Issue: 159(5)

Published: Aug. 1, 2023

DeePMD-kit is a powerful open-source software package that facilitates molecular dynamics simulations using machine learning potentials known as Deep Potential (DP) models. This package, which was released in 2017, has been widely used the fields of physics, chemistry, biology, and material science for studying atomistic systems. The current version offers numerous advanced features, such DeepPot-SE, attention-based hybrid descriptors, ability to fit tensile properties, type embedding, model deviation, DP-range correction, DP long range, graphics processing unit support customized operators, compression, non-von Neumann dynamics, improved usability, including documentation, compiled binary packages, graphical user interfaces, application programming interfaces. article presents an overview major highlighting its features technical details. Additionally, this comprehensive procedure conducting representative application, benchmarks accuracy efficiency different models, discusses ongoing developments.

Language: Английский

---

MLatom 3: A Platform for Machine Learning-Enhanced Computational Chemistry Simulations and Workflows

Journal of Chemical Theory and Computation, Journal Year: 2024, Volume and Issue: 20(3), P. 1193 - 1213

Published: Jan. 25, 2024

Machine learning (ML) is increasingly becoming a common tool in computational chemistry. At the same time, rapid development of ML methods requires flexible software framework for designing custom workflows. MLatom 3 program package designed to leverage power enhance typical chemistry simulations and create complex This open-source provides plenty choice users who can run with command-line options, input files, or scripts using as Python package, both on their computers online XACS cloud computing service at XACScloud.com. Computational chemists calculate energies thermochemical properties, optimize geometries, molecular quantum dynamics, simulate (ro)vibrational, one-photon UV/vis absorption, two-photon absorption spectra ML, mechanical, combined models. The choose from an extensive library containing pretrained models mechanical approximations such AIQM1 approaching coupled-cluster accuracy. developers build own various algorithms. great flexibility largely due use interfaces many state-of-the-art packages libraries.

Language: Английский

---

TorchMD-Net 2.0: Fast Neural Network Potentials for Molecular Simulations

Journal of Chemical Theory and Computation, Journal Year: 2024, Volume and Issue: 20(10), P. 4076 - 4087

Published: May 14, 2024

Achieving a balance between computational speed, prediction accuracy, and universal applicability in molecular simulations has been persistent challenge. This paper presents substantial advancements TorchMD-Net software, pivotal step forward the shift from conventional force fields to neural network-based potentials. The evolution of into more comprehensive versatile framework is highlighted, incorporating cutting-edge architectures such as TensorNet. transformation achieved through modular design approach, encouraging customized applications within scientific community. most notable enhancement significant improvement efficiency, achieving very remarkable acceleration computation energy forces for TensorNet models, with performance gains ranging 2× 10× over previous, nonoptimized, iterations. Other enhancements include highly optimized neighbor search algorithms that support periodic boundary conditions smooth integration existing dynamics frameworks. Additionally, updated version introduces capability integrate physical priors, further enriching its application spectrum utility research. software available at https://github.com/torchmd/torchmd-net.

Language: Английский

---

Mechanism of Charge Transport in Lithium Thiophosphate

Chemistry of Materials, Journal Year: 2024, Volume and Issue: 36(3), P. 1482 - 1496

Published: Feb. 5, 2024

Lithium ortho-thiophosphate (Li3PS4) has emerged as a promising candidate for solid-state electrolyte batteries, thanks to its highly conductive phases, cheap components, and large electrochemical stability range. Nonetheless, the microscopic mechanisms of Li-ion transport in Li3PS4 are far from being fully understood, role PS4 dynamics charge still controversial. In this work, we build machine learning potentials targeting state-of-the-art DFT references (PBEsol, r2SCAN, PBE0) tackle problem all known phases (α, β, γ), system sizes time scales. We discuss physical origin observed superionic behavior Li3PS4: activation flipping drives structural transition phase, characterized by an increase Li-site availability drastic reduction energy diffusion. also rule out any paddle-wheel effects tetrahedra phases─previously claimed enhance diffusion─due orders-of-magnitude difference between rate flips hops at temperatures below melting. finally elucidate interionic dynamical correlations transport, highlighting failure Nernst–Einstein approximation estimate electrical conductivity. Our results show strong dependence on target reference, with PBE0 yielding best quantitative agreement experimental measurements not only electronic band gap but conductivity β- α-Li3PS4.

Language: Английский

---

In Silico Chemical Experiments in the Age of AI: From Quantum Chemistry to Machine Learning and Back

Advanced Materials, Journal Year: 2024, Volume and Issue: 36(30)

Published: May 25, 2024

Abstract Computational chemistry is an indispensable tool for understanding molecules and predicting chemical properties. However, traditional computational methods face significant challenges due to the difficulty of solving Schrödinger equations increasing cost with size molecular system. In response, there has been a surge interest in leveraging artificial intelligence (AI) machine learning (ML) techniques silico experiments. Integrating AI ML into increases scalability speed exploration space. remain, particularly regarding reproducibility transferability models. This review highlights evolution from, complementing, or replacing energy property predictions. Starting from models trained entirely on numerical data, journey set forth toward ideal model incorporating physical laws quantum mechanics. paper also reviews existing their intertwining, outlines roadmap future research, identifies areas improvement innovation. Ultimately, goal develop architectures capable accurate transferable solutions equation, thereby revolutionizing experiments within materials science.

Language: Английский

---

LASP to the Future of Atomic Simulation: Intelligence and Automation

Precision Chemistry, Journal Year: 2024, Volume and Issue: 2(12), P. 612 - 627

Published: Sept. 14, 2024

Atomic simulations aim to understand and predict complex physical phenomena, the success of which relies largely on accuracy potential energy surface description efficiency capture important rare events. LASP software (large-scale atomic simulation with a Neural Network Potential), released in 2018, incorporates key ingredients fulfill ultimate goal by combining advanced neural network potentials efficient global optimization methods. This review introduces recent development along two main streams, namely, higher intelligence more automation, solve material reaction problems. The latest version (LASP 3.7) features many-body function corrected (G-MBNN) improve PES low cost, achieves linear scaling for large-scale simulations. functionalities are updated incorporate (i) ASOP ML-interface methods finding interface structures under grand canonic conditions; (ii) ML-TS MMLPS identify lowest pathway. With these powerful functionalities, now serves as an intelligent data generator create computational databases end users. We exemplify database construction zeolite metal-ligand properties new catalyst design.

Language: Английский

---

Towards symbolic XAI – explanation through human understandable logical relationships between features

Information Fusion, Journal Year: 2025, Volume and Issue: unknown, P. 102923 - 102923

Published: Jan. 1, 2025

Language: Английский

---

Analyzing Atomic Interactions in Molecules as Learned by Neural Networks

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

Published: Jan. 10, 2025

While machine learning (ML) models have been able to achieve unprecedented accuracies across various prediction tasks in quantum chemistry, it is now apparent that accuracy on a test set alone not guarantee for robust chemical modeling such as stable molecular dynamics (MD). To go beyond accuracy, we use explainable artificial intelligence (XAI) techniques develop general analysis framework atomic interactions and apply the SchNet PaiNN neural network models. We compare these with of fundamental principles understand how well learned underlying physicochemical concepts from data. focus strength different species, predictions intensive extensive properties are made, analyze decay many-body nature interatomic distance. Models deviate too far known physical produce unstable MD trajectories, even when they very high energy force accuracy. also suggest further improvements ML architectures better account polynomial interactions.

Language: Английский

---

Machine Learning Interatomic Potentials for Reactive Hydrogen Dynamics at Metal Surfaces Based on Iterative Refinement of Reaction Probabilities

The Journal of Physical Chemistry C, Journal Year: 2023, Volume and Issue: 127(50), P. 24168 - 24182

Published: Dec. 4, 2023

The reactive chemistry of molecular hydrogen at surfaces, notably dissociative sticking and evolution, plays a crucial role in energy storage fuel cells. Theoretical studies can help to decipher underlying mechanisms reaction design, but studying dynamics surfaces is computationally challenging due the complex electronic structure interfaces high sensitivity barriers. In addition, ab initio dynamics, based on density functional theory, too demanding accurately predict or desorption probabilities, as it requires averaging over tens thousands initial conditions. High-dimensional machine learning-based interatomic potentials are starting be more commonly used gas-surface yet robust approaches generate reliable training data assess how model uncertainty affects prediction dynamic observables not well established. Here, we employ ensemble learning adaptively while assessing performance with full quantification (UQ) for probabilities scattering different copper facets. We use this approach investigate two message-passing neural networks, SchNet PaiNN. Ensemble-based UQ iterative refinement allow us expose shortcomings invariant pairwise-distance-based feature representation dynamics.

Language: Английский

---

Benchmarking of machine learning interatomic potentials for reactive hydrogen dynamics at metal surfaces

Machine Learning Science and Technology, Journal Year: 2024, Volume and Issue: 5(3), P. 030501 - 030501

Published: July 3, 2024

Abstract Simulations of chemical reaction probabilities in gas surface dynamics require the calculation ensemble averages over many tens thousands events to predict dynamical observables that can be compared experiments. At same time, energy landscapes need accurately mapped, as small errors barriers lead large deviations probabilities. This brings a particularly interesting challenge for machine learning interatomic potentials, which are becoming well-established tools accelerate molecular simulations. We compare state-of-the-art potentials with particular focus on their inference performance CPUs and suitability high throughput simulation reactive chemistry at surfaces. The considered models include polarizable atom interaction neural networks (PaiNN), recursively embedded (REANN), MACE equivariant graph network, atomic cluster expansion (ACE). applied dataset hydrogen scattering low-index facets copper. All assessed accuracy, time-to-solution, ability simulate sticking function rovibrational initial state kinetic incidence molecule. REANN provide best balance between accuracy time-to-solution current gas-surface dynamics. PaiNN features causes significant losses computational efficiency. ACE fastest however, trained existing were not able achieve sufficiently accurate predictions all cases.

Language: Английский

---