Single-model uncertainty quantification in neural network potentials does not consistently outperform model ensembles

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

Published: Dec. 16, 2023

Language: Английский

---

Deep ensembles vs committees for uncertainty estimation in neural-network force fields: Comparison and application to active learning

The Journal of Chemical Physics, Journal Year: 2023, Volume and Issue: 158(20)

Published: May 22, 2023

A reliable uncertainty estimator is a key ingredient in the successful use of machine-learning force fields for predictive calculations. Important considerations are correlation with error, overhead during training and inference, efficient workflows to systematically improve field. However, case neural-network fields, simple committees often only option considered due their easy implementation. Here, we present generalization deep-ensemble design based on multiheaded neural networks heteroscedastic loss. It can efficiently deal uncertainties both energy forces take sources aleatoric affecting data into account. We compare metrics deep ensembles, committees, bootstrap-aggregation ensembles using an ionic liquid perovskite surface. demonstrate adversarial approach active learning progressively refine fields. That workflow realistically possible thanks exceptionally fast enabled by residual nonlinear learned optimizer.

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

---

Transferability and Accuracy of Ionic Liquid Simulations with Equivariant Machine Learning Interatomic Potentials

The Journal of Physical Chemistry Letters, Journal Year: 2024, Volume and Issue: 15(30), P. 7539 - 7547

Published: July 18, 2024

Ionic liquids (ILs) are an exciting class of electrolytes finding applications in many areas from energy storage to solvents, where they have been touted as "designer solvents" can be mixed precisely tailor the physiochemical properties. As using machine learning interatomic potentials (MLIPs) simulate ILs is still relatively unexplored, several questions need answered see if MLIPs transformative for ILs. Since often not pure, but either together or contain additives, we first demonstrate that a MLIP trained compositionally transferable; i.e., applied mixtures ions directly on, while only being on few same ions. We also investigated accuracy novel IL, which experimentally synthesize and characterize. Our ∼200 DFT frames reasonable agreement with our experiments DFT.

Language: Английский

---

Accurate machine learning force fields via experimental and simulation data fusion

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

Published: April 5, 2024

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

---

Transferable Water Potentials Using Equivariant Neural Networks

The Journal of Physical Chemistry Letters, Journal Year: 2024, Volume and Issue: 15(14), P. 3740 - 3747

Published: March 28, 2024

Machine learning interatomic potentials (MLIPs) have emerged as a technique that promises quantum theory accuracy for reduced cost. It has been proposed [J. Chem. Phys. 2023, 158, 084111] MLIPs trained on solely liquid water data cannot accurately transfer to the vapor-liquid equilibrium while recovering many-body decomposition (MBD) analysis of gas-phase clusters. This suggests do not directly learn physically correct interactions molecules, limiting transferability. In this work, we show using equivariant architecture and 3200 structures reproduces liquid-phase properties (e.g., density within 0.003 g/cm3 between 230 365 K), up 550 K, MBD cluster six-body interactions, relative energy vibrational states ice phases. We developed allow transferability arbitrary phases remain stable in nanosecond long simulations.

Language: Английский

---

Uncertainty quantification by direct propagation of shallow ensembles

Machine Learning Science and Technology, Journal Year: 2024, Volume and Issue: 5(3), P. 035006 - 035006

Published: June 17, 2024

Abstract Statistical learning algorithms provide a generally-applicable framework to sidestep time-consuming experiments, or accurate physics-based modeling, but they introduce further source of error on top the intrinsic limitations experimental theoretical setup. Uncertainty estimation is essential quantify this error, and make application data-centric approaches more trustworthy. To ensure that uncertainty quantification used widely, one should aim for are accurate, also easy implement apply. In particular, including an existing architecture be straightforward, add minimal computational overhead. Furthermore, it manipulate combine multiple machine-learning predictions, propagating over modeling steps. We compare several well-established frameworks against these requirements, propose practical approach, which we dub direct propagation shallow ensembles, provides good compromise between ease use accuracy. present benchmarks generic datasets, in-depth study applications field atomistic machine chemistry materials. These examples underscore importance using formulation allows errors without making strong assumptions correlations different predictions model.

Language: Английский

---

Applications of machine‐learning interatomic potentials for modeling ceramics, glass, and electrolytes: A review

Journal of the American Ceramic Society, Journal Year: 2024, Volume and Issue: unknown

Published: June 9, 2024

Abstract The emergence of artificial intelligence has provided efficient methodologies to pursue innovative findings in material science. Over the past two decades, machine‐learning potential (MLP) emerged as an alternative technology density functional theory (DFT) and classical molecular dynamics (CMD) simulations for computational modeling materials estimation their properties. MLP offers more computation compared DFT, while providing higher accuracy CMD. This enables us conduct realistic using models with atoms longer simulation times. Indeed, number research studies utilizing MLPs significantly increased since 2015, covering a broad range structures, ranging from simple complex, well various chemical physical phenomena. As result, there are high expectations further applications field science industrial development. review aims summarize applications, particularly ceramics glass science, fundamental theories facilitate future progress utilization. Finally, we provide summary discuss perspectives on next challenges development application MLPs.

Language: Английский

---

Scaling the Leading Accuracy of Deep Equivariant Models to Biomolecular Simulations of Realistic Size

Published: Nov. 11, 2023

This work brings the leading accuracy, sample efficiency, and robustness of deep equivariant neural networks to extreme computational scale. is achieved through a combination innovative model architecture, massive parallelization, models implementations optimized for efficient GPU utilization. The resulting Allegro architecture bridges accuracy-speed tradeoff atomistic simulations enables description dynamics in structures unprecedented complexity at quantum fidelity. To illustrate scalability Allegro, we perform nanoseconds-long stable protein scale up 44-million atom structure complete, all-atom, explicitly solvated HIV capsid on Perlmutter supercomputer. We demonstrate excellent strong scaling 100 million atoms 70% weak 5120 A100 GPUs.

Language: Английский

---