Scientific discovery in the age of artificial intelligence

Nature, Journal Year: 2023, Volume and Issue: 620(7972), P. 47 - 60

Published: Aug. 2, 2023

Language: Английский

---

CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling

Nature Machine Intelligence, Journal Year: 2023, Volume and Issue: 5(9), P. 1031 - 1041

Published: Sept. 14, 2023

Abstract Large-scale simulations with complex electron interactions remain one of the greatest challenges for atomistic modelling. Although classical force fields often fail to describe coupling between electronic states and ionic rearrangements, more accurate ab initio molecular dynamics suffers from computational complexity that prevents long-time large-scale simulations, which are essential study technologically relevant phenomena. Here we present Crystal Hamiltonian Graph Neural Network (CHGNet), a graph neural network-based machine-learning interatomic potential (MLIP) models universal energy surface. CHGNet is pretrained on energies, forces, stresses magnetic moments Materials Project Trajectory Dataset, consists over 10 years density functional theory calculations than 1.5 million inorganic structures. The explicit inclusion enables learn accurately represent orbital occupancy electrons, enhancing its capability both atomic degrees freedom. We demonstrate several applications in solid-state materials, including charge-informed Li x MnO 2 , finite temperature phase diagram FePO 4 diffusion garnet conductors. highlight significance charge information capturing appropriate chemistry provide insights into systems additional freedom cannot be observed by previous MLIPs.

Language: Английский

---

General framework for E(3)-equivariant neural network representation of density functional theory Hamiltonian

Nature Communications, Journal Year: 2023, Volume and Issue: 14(1)

Published: May 18, 2023

Combination of deep learning and ab initio calculation has shown great promise in revolutionizing future scientific research, but how to design neural network models incorporating a priori knowledge symmetry requirements is key challenging subject. Here we propose an E(3)-equivariant deep-learning framework represent density functional theory (DFT) Hamiltonian as function material structure, which can naturally preserve the Euclidean even presence spin-orbit coupling. Our DeepH-E3 method enables very efficient electronic-structure at accuracy by from DFT data small-sized structures, making routine study large-scale supercells ($> 10^4$ atoms) feasible. Remarkably, reach sub-meV prediction high training efficiency, showing state-of-the-art performance our experiments. The work not only general significance development, also creates new opportunities for materials such building Moir\'e-twisted database.

Language: Английский

---

Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing

Nature Communications, Journal Year: 2024, Volume and Issue: 15(1)

Published: Jan. 5, 2024

Abstract Geometric deep learning has been revolutionizing the molecular modeling field. Despite state-of-the-art neural network models are approaching ab initio accuracy for property prediction, their applications, such as drug discovery and dynamics (MD) simulation, have hindered by insufficient utilization of geometric information high computational costs. Here we propose an equivariant geometry-enhanced graph called ViSNet, which elegantly extracts features efficiently structures with low Our proposed ViSNet outperforms approaches on multiple MD benchmarks, including MD17, revised MD17 MD22, achieves excellent chemical prediction QM9 Molecule3D datasets. Furthermore, through a series simulations case studies, can explore conformational space provide reasonable interpretability to map representations structures.

Language: Английский

---

Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations

iScience, Journal Year: 2024, Volume and Issue: 27(5), P. 109673 - 109673

Published: April 4, 2024

Machine learning interatomic potential (MLIP) overcomes the challenges of high computational costs in density-functional theory and relatively low accuracy classical large-scale molecular dynamics, facilitating more efficient precise simulations materials research design. In this review, current state four essential stages MLIP is discussed, including data generation methods, material structure descriptors, six unique machine algorithms, available software. Furthermore, applications various fields are investigated, notably phase-change memory materials, searching, properties predicting, pre-trained universal models. Eventually, future perspectives, consisting standard datasets, transferability, generalization, trade-off between complexity MLIPs, reported.

Language: Английский

---

Crash testing machine learning force fields for molecules, materials, and interfaces: model analysis in the TEA Challenge 2023

Chemical Science, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 1, 2025

Assessing the performance of modern machine learning force fields across diverse chemical systems to identify their strengths and limitations within TEA Challenge 2023.

Language: Английский

---

Informing geometric deep learning with electronic interactions to accelerate quantum chemistry

Proceedings of the National Academy of Sciences, Journal Year: 2022, Volume and Issue: 119(31)

Published: July 28, 2022

Predicting electronic energies, densities, and related chemical properties can facilitate the discovery of novel catalysts, medicines, battery materials. However, existing machine learning techniques are challenged by scarcity training data when exploring unknown spaces. We overcome this barrier systematically incorporating knowledge molecular structure into deep learning. By developing a physics-inspired equivariant neural network, we introduce method to learn representations based on interactions among atomic orbitals. Our method, OrbNet-Equi, leverages efficient tight-binding simulations learned mappings recover high-fidelity physical quantities. OrbNet-Equi accurately models wide spectrum target while being several orders magnitude faster than density functional theory. Despite only using samples collected from readily available small-molecule libraries, outperforms traditional semiempirical learning-based methods comprehensive downstream benchmarks that encompass diverse main-group processes. also describes in challenging charge-transfer complexes open-shell systems. anticipate strategy presented here will help expand opportunities for studies chemistry materials science, where acquisition experimental or reference is costly.

Language: Английский

---

Computational Studies of Aflatoxin B1 (AFB1): A Review

Toxins, Journal Year: 2023, Volume and Issue: 15(2), P. 135 - 135

Published: Feb. 7, 2023

Aflatoxin B1 (AFB1) exhibits the most potent mutagenic and carcinogenic activity among aflatoxins. For this reason, AFB1 is recognized as a human group 1 carcinogen by International Agency of Research on Cancer. Consequently, it essential to determine its properties behavior in different chemical systems. The can be explored using computational chemistry, which has been employed complementarily experimental investigations. present review includes silico studies (semiempirical, Hartree–Fock, DFT, molecular docking, dynamics) conducted from first study 1974 (2022). This work was performed, considering following groups: (a) (structural, energy, solvent effects, ground excited state, atomic charges, others); (b) theoretical investigations (degradation, quantification, reactivity, (c) interactions with inorganic compounds (Ag+, Zn2+, Mg2+); (d) environmentally (clays); (e) biological (DNA, enzymes, cyclodextrins, glucans, others). Accordingly, work, we provide stakeholder knowledge toxicity types AFB1-derivatives, structure–activity relationships manifested bonds between DNA or proteins, strategies that have quantify, detect, eliminate molecule.

Language: Английский

---

Fast uncertainty estimates in deep learning interatomic potentials

The Journal of Chemical Physics, Journal Year: 2023, Volume and Issue: 158(16)

Published: April 27, 2023

Deep learning has emerged as a promising paradigm to give access highly accurate predictions of molecular and material properties. A common short-coming shared by current approaches, however, is that neural networks only point estimates their do not come with predictive uncertainties associated these estimates. Existing uncertainty quantification efforts have primarily leveraged the standard deviation across an ensemble independently trained networks. This incurs large computational overhead in both training prediction, resulting order-of-magnitude more expensive predictions. Here, we propose method estimate based on single network without need for ensemble. allows us obtain virtually no additional over inference. We demonstrate quality matches those obtained from deep ensembles. further examine our methods ensembles configuration space test system compare potential energy surface. Finally, study efficacy active setting find results match ensemble-based strategy at reduced cost.

Language: Английский

---

Exploring protein–ligand binding affinity prediction with electron density-based geometric deep learning

RSC Advances, Journal Year: 2024, Volume and Issue: 14(7), P. 4492 - 4502

Published: Jan. 1, 2024

A deep learning approach centered on electron density is suggested for predicting the binding affility between proteins and ligands. The thoroughly assessed using various pertinent benchmarks.

Language: Английский

---