Constructing Accurate and Efficient General-Purpose Atomistic Machine Learning Model with Transferable Accuracy for Quantum Chemistry

Journal of Chemical Theory and Computation, Journal Year: 2024, Volume and Issue: 20(21), P. 9500 - 9511

Published: Oct. 31, 2024

Density functional theory (DFT) has been a cornerstone in computational science, providing powerful insights into structure-property relationships for molecules and materials through first-principles quantum-mechanical (QM) calculations. However, the advent of atomistic machine learning (ML) is reshaping landscape by enabling large-scale dynamics simulations high-throughput screening at DFT-equivalent accuracy with drastically reduced cost. Yet, development general-purpose ML models as surrogates QM calculations faces several challenges, particularly terms model capacity, data efficiency, transferability across chemically diverse systems. This work introduces novel extension polarizable atom interaction neural network (namely, XPaiNN) to address these challenges. Two distinct training strategies have employed, one direct-learning other Δ-ML on top semiempirical method. These methodologies implemented within same framework, allowing detailed comparison their results. The XPaiNN models, particular using Δ-ML, not only demonstrate competitive performance standard benchmarks, but also effectiveness against methods comprehensive downstream tasks, including noncovalent interactions, reaction energetics, barrier heights, geometry optimization thermodynamics, etc. represents significant step forward pursuit accurate efficient general-purpose, capable handling complex chemical systems transferable accuracy.

Language: Английский

---

Benchmark of general-purpose machine learning-based quantum mechanical method AIQM1 on reaction barrier heights

The Journal of Chemical Physics, Journal Year: 2023, Volume and Issue: 158(7)

Published: Jan. 30, 2023

Artificial intelligence-enhanced quantum mechanical method 1 (AIQM1) is a general-purpose that was shown to achieve high accuracy for many applications with speed close its baseline semiempirical (SQM) ODM2*. Here, we evaluate the hitherto unknown performance of out-of-the-box AIQM1 without any refitting reaction barrier heights on eight datasets, including total ∼24 thousand reactions. This evaluation shows AIQM1's strongly depends type transition state and ranges from excellent rotation barriers poor for, e.g., pericyclic clearly outperforms ODM2* and, even more so, popular universal potential, ANI-1ccx. Overall, however, largely remains similar SQM methods (and B3LYP/6-31G* most types) suggesting it desirable focus improving in future. We also show built-in uncertainty quantification helps identifying confident predictions. The predictions approaching level density functional theory types. Encouragingly, rather robust optimizations, reactions struggles most. Single-point calculations high-level AIQM1-optimized geometries can be used significantly improve heights, which cannot said method.

Language: Английский

---

Toward DMC Accuracy Across Chemical Space with Scalable Δ-QML

Journal of Chemical Theory and Computation, Journal Year: 2023, Volume and Issue: 19(6), P. 1711 - 1721

Published: March 1, 2023

In the past decade, quantum diffusion Monte Carlo (DMC) has been demonstrated to successfully predict energetics and properties of a wide range molecules solids by numerically solving electronic many-body Schrödinger equation. With O(N3) scaling with number electrons N, DMC potential be reference method for larger systems that are not accessible more traditional methods such as CCSD(T). Assessing accuracy smaller becomes stepping stone in making systems. We show when coupled machine learning (QML)-based surrogate methods, computational burden can alleviated (QMC) shows clear undergird formation high-quality descriptions across chemical space. discuss three crucial approximations necessary accomplish this: fixed-node approximation, universal accurate references bond dissociation energies, scalable minimal amons-set-based QML (AQML) models. Numerical evidence presented includes converged results over 1000 small organic up five heavy atoms used amons 50 medium-sized nine validate AQML predictions. collected Δ-AQML models suggests already modestly sized QMC training data sets suffice total energies near throughout

Language: Английский

---

Acceleration of Graph Neural Network-Based Prediction Models in Chemistry via Co-Design Optimization on Intelligence Processing Units

Journal of Chemical Information and Modeling, Journal Year: 2024, Volume and Issue: 64(5), P. 1568 - 1580

Published: Feb. 21, 2024

Atomic structure prediction and associated property calculations are the bedrock of chemical physics. Since high-fidelity ab initio modeling techniques for computing properties can be prohibitively expensive, this motivates development machine-learning (ML) models that make these predictions more efficiently. Training graph neural networks over large atomistic databases introduces unique computational challenges, such as need to process millions small graphs with variable size support communication patterns distinct from learning graphs, social networks. We demonstrate a novel hardware–software codesign approach scale up training (GNN) prediction. First, eliminate redundant computation memory alternative padding improve throughput via minimizing communication, we formulate effective coalescing batches variable-size bin packing problem introduce hardware-agnostic algorithm pack batches. In addition, propose hardware-specific optimizations, including planner vectorization gather-scatter operations targeted Graphcore's Intelligence Processing Unit (IPU), well model-specific optimizations merged collectives optimized softplus. Putting all together, effectiveness proposed by providing an implementation well-established GNN on Graphcore IPUs. evaluate performance multiple varying degrees counts, sizes, sparsity. reduce time GNNs their 1.5× compared baseline model Additionally, compare our IPU Nvidia GPU-based show IPUs run 1.8× faster average execution GPUs.

Language: Английский

---

Universal and Updatable Artificial Intelligence-Enhanced Quantum Chemical Foundational Models

Published: June 26, 2024

Quantum chemical methods developed since 1927 are instrumental in simulations but human expertise has been still essential choosing a suitable method. Here we introduce paradigm shift to universal and updatable artificial intelligence-enhanced quantum mechanical (UAIQM) foundational models with an online platform auto-selecting the best accuracy for given system, available time, moderate computational resources (see https://xacs.xmu.edu.cn/docs/mlatom/tutorial_uaiqm.html instructions). The hosts growing library of state-of-the-art UAIQM calibrated uncertainties provides mechanism improving continuously more usage. We demonstrate how can be used massive accurate within hours on commodity hardware which would take days or weeks high-performance computing centers less workhorse methods. also show that sets new standard infrared spectra, reaction barriers, energetics whose predictions have far-reaching consequences molecular simulations.

Language: Английский

---