Refining potential energy surface through dynamical properties via differentiable molecular simulation

Nature Communications, Journal Year: 2025, Volume and Issue: 16(1)

Published: Jan. 18, 2025

Recently, machine learning potential (MLP) largely enhances the reliability of molecular dynamics, but its accuracy is limited by underlying ab initio methods. A viable approach to overcome this limitation refine from experimental data, which now can be done efficiently using modern automatic differentiation technique. However, refinement mostly performed thermodynamic properties, leaving most accessible and informative dynamical data (like spectroscopy) unexploited. In work, through a comprehensive application adjoint gradient truncation methods, we show that both memory explosion issues circumvented in many situations, so property well-behaved. Consequently, transport coefficients spectroscopic used improve density functional theory based MLP towards higher accuracy. Essentially, work contributes solution inverse problem spectroscopy extracting microscopic interactions vibrational data.

Language: Английский

---

A simple approach to rotationally invariant machine learning of a vector quantity

The Journal of Chemical Physics, Journal Year: 2024, Volume and Issue: 161(17)

Published: Nov. 1, 2024

Unlike with the energy, which is a scalar property, machine learning (ML) prediction of vector or tensor properties poses additional challenge achieving proper invariance (covariance) respect to molecular rotation. For energy gradients needed in dynamics (MD), this symmetry automatically fulfilled when taking analytic derivative invariant (using properly descriptors). However, if cannot be obtained by differentiation, other appropriate methods should applied retain covariance. Several approaches have been suggested treat issue. nonadiabatic couplings and polarizabilities, for example, it was possible construct virtual quantities from above tensorial are differentiation thus guarantee Another solution build rotational equivariance into design neural network employed model. Here, we propose simpler alternative technique, does not require construction auxiliary application special equivariant ML techniques. We suggest three-step approach, using inertia. In first step, molecule rotated eigenvectors its principal axes. second procedure predicts property relative orientation, based on training set where all were same coordinate system. As third remains transform estimate back original orientation. This rotate–predict–rotate (RPR) covariance trivially extensible also tensors such as polarizability. The RPR has an advantage that accurate models can trained very fast thousands configurations, might beneficial many sets required (e.g., active learning). implemented MLatom Newton-X programs MD, performed assessment dipole moment along MD trajectories 1,2-dichloroethane.

Language: Английский

---

Accurate Modeling of the Potential Energy Surface of Molecular Clusters Boosted by Neural Networks

Environmental Science Advances, Journal Year: 2024, Volume and Issue: 3(10), P. 1438 - 1451

Published: Jan. 1, 2024

We present the application of machine learning methods to alleviate computational cost quantum chemistry calculations required for modeling atmospheric molecular clusters.

Language: Английский

---

The Analysis of Vibrational Spectra: Past, Present and Future

ChemPlusChem, Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 11, 2024

Vibrational spectroscopy can be said to have started with the seminal work of Coblentz in 1900s, who recorded first recognisable infrared spectra. Today, vibrational is ubiquitous and there are many ways measure a spectrum. But this usually only step, almost always need assign resulting spectra: "what property system results feature at energy"? How question has been answered changed over last century, as our understanding fundamental physics matter evolved. In Perspective, I will present my view how analysis spectra evolved time. The article divided into three sections: past, future. "past" section consists very brief history spectroscopy. "present" centered around ab initio studies, particularly density functional theory (DFT) describe become routine. For "future", extrapolate current trends also speculate what might come next.

Language: Английский

---