A machine-learning framework for accelerating spin-lattice relaxation simulations - Applying advanced ML workflow to predict vibrations and spinphonon coupling

Aug 4, 2026

A machine-learning framework accelerates spin-lattice relaxation simulations by predicting molecular vibrations and spinphonon coupling coefficients with semito-full quantitative agreement to ab initio methods. The approach targets open-shell coordination compounds with long relaxation times and reduces the computational cost by about 80% compared with expensive phonon and derivative calculations. The workflow also extends to molecular dynamics simulations, enabling spin relaxation studies beyond simple equilibrium harmonic thermal baths.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages34

Reading Time1 h 26 min

AccessFree to download

Introduction

Computational bottlenecks in ab initio spin relaxation simulations

ML-based force fields and phonon prediction

Goal of this work: an ML workflow for spin relaxation

A machine-learning framework for accelerating spin-lattice relaxation simulations arXiv :2410 .08912v1 [ cond-mat .mtrl-sci ] 11 Oct 2024 Valerio Briganti and Alessandro Lunghi ∗ School of Physics, AMBER and CRANN Institute, Trinity College, Dublin 2, Ireland Molecular and lattice vibrations are able to couple to the spin of electrons and lead to their relaxation and decoherence. Ab initio simulations have played a fundamental role in shaping our understanding of this process but further progress is hindered by their high computational cost. Here we present an accelerated computational framework based on machine-learning models for the prediction of molecular vibrations and spinphonon coupling coefficients. We apply this method to three open-shell coordination compounds exhibiting long relaxation times and show that this approach achieves semito-full quantitative agreement with ab initio methods reducing the computational cost by about 80% . Moreover, we show that this framework naturally extends to molecular dynamics simulations, paving the way to the study of spin relaxation in condensed matter beyond simple equilibrium harmonic thermal baths. I. INTRODUCTION The study of spin dynamics of coordination compounds is central to the most recent advancements in different fields such as information storage, sensing, information processing, spintronics and paramagnetic resonance for biosystems [1–3] . Understanding how spin relaxes towards equilibrium in molecules represents a longstanding puzzle whose solution would lead to a better interpretation of experiments as well as powering computational design of new magnetic materials [4] . Ab initio spin dynamics simulations based on open quantum systems theory, density functional theory (DFT) and multiconfigurational quantum chemistry techniques are nowadays considered a gold standard tool to perform spin relaxation simulations [4] . This methodology has now been successfully applied to a large number of magnetic systems, including spin-1/2[5–7], singlemolecule magnets[8–11], solid-state defects[12, 13], and surface-adsorbed atoms[14] . However, these simulations require the numerical calculation of phonons for hundreds of atoms and derivatives with multi-reference ab initio methods, making predictions extremely expensive[4] . Moreover, whilst full ab initio calculation of a few systems at the time is possible through modern-day highperformance computing resources, the discovery of new materials with long relaxation times will inevitably require the characterization of many compounds[15, 16], making a brute force ab initio approach unsuitable. Machine learning (ML) has gained increasing popularity in science over the last years as a tool to overcome known computational bottlenecks of ab initio methods or to extract hidden patterns in big data. Relevant to this work is its application in the generation of force fields (FF) from a dataset of ab initio data, referred to as a training set. A MLFF is specified by a parametric function that links atomic coordinates and the total energy of a system. The MLFF’s parameters are de- ∗ [lunghia@tcd.ie](lunghia@tcd.ie) termined by minimization of the differences between the model’s predictions and the full quantum mechanical calculations. Once the parameters of the model are determined, i.e. the model is trained, the MLFF can rapidly output the energy and atomic forces for any molecular configuration, thus avoiding the expensive step of solving the Schr¨odinger equation. This field has extended considerably in the last years and a plethora of different MLFFs are now available [17–30] and capable of reaching chemical accuracy in their predictions of several material properties [31–34] . Importantly, MLFFs have been used to predict phonon spectra of a large variety of compounds such as molecular crystals [35], organic molecules [36] and inorganic solid compounds [37], and new schemes are currently under study to achieve high trans

FAQs

Why are ab initio spin relaxation simulations computationally expensive?

They require numerical phonon calculations for hundreds of atoms and derivatives using multi-reference ab initio methods, which makes predictions extremely costly.

What does the proposed machine-learning framework predict?

It predicts molecular vibrations and spinphonon coupling coefficients, using ML models trained on ab initio data.

How much computational cost does the ML framework save, and how is agreement validated?

It reduces computational cost by about 80% while achieving semito-full quantitative agreement with ab initio methods, including minimal accuracy loss when using ML for both vibrations and coupling.

Aug 4, 2026

Upload by Valentina