Machine Learned Force Fields: Fundamentals, its reach, and challenges

TL;DR

The chapter surveys machine-learned force fields (MLFFs) as a route to achieve near-quantum accuracy in atomistic simulations with orders-of-magnitude speedups over ab initio methods. It contrasts neural-network-based approaches (e.g., SchNet and its modern equivariant successors) with kernel-based frameworks (GDML, GAP, sGDML), detailing how end-to-end learned representations and physics-inspired kernels are used to model potential energy surfaces and forces. Key contributions include a detailed exposition of SchNet’s continuous-filter convolutions and atom-type embeddings, as well as the GDML framework that directly learns forces via a Hessian kernel and can be symmetrized for enhanced data efficiency. The discussion highlights current challenges—data requirements, transferability, and uncertainty quantification—and argues that MLFFs are reshaping atomistic modeling by enabling accurate, scalable simulations across chemistry and materials science, with broad implications for discovery and design.

Abstract

Highly accurate force fields are a mandatory requirement to generate predictive simulations. In this regard, Machine Learning Force Fields (MLFFs) have emerged as a revolutionary approach in computational chemistry and materials science, combining the accuracy of quantum mechanical methods with computational efficiency orders of magnitude superior to ab-initio methods. This chapter provides an introduction of the fundamentals of learning and how it is applied to construct MLFFs, detailing key methodologies such as neural network potentials and kernel-based models. Emphasis is placed on the construction of SchNet model, as one of the most elemental neural network-based force fields that are nowadays the basis of modern architectures. Additionally, the GDML framework is described in detail as an example of how the elegant formulation of kernel methods can be used to construct mathematically robust and physics-inspired MLFFs. The ongoing advancements in MLFF development continue to expand their applicability, enabling precise simulations of large and complex systems that were previously beyond reach. This chapter concludes by highlighting the transformative impact of MLFFs on scientific research, underscoring their role in driving future discoveries in the fields of chemistry, physics, and materials science.

Figures (9)

• Figure 1: A) Machine learning starts from quantum calculations to learn material behavior through analytical potentials. B) Taking atomic coordinates, among other quantities as inputs, neural networks compute energies and learn atomic force fields as outputs. C) Common sub-processes for machine learning algorithms in material research. Mainly algorithms start with the codification of material configurations and continue with the learning through convolutional neural networks. And finally, multilayer perceptrons compute properties.
• Figure 2: A) Database types: Image classification, force field learning, and materials properties. B) Types of Learning. Supervised learning and unsupervised learning.
• Figure 3: What does Learning mean? A) Model construction. B) Models training and generalization.
• Figure 4: Examples of activation functions. A) Threshold (or step) function. B) Sigmoid Function. C) ReLU function. D) SiLU function.
• Figure 5: A) Graphic view of a Neuron Model. B) Example of classification of two classes. C), D) and E) shows examples of neural network architectures. C) Multi-layer perceptron. D) Deep Recurrent Neural Network. E) Convolutional Neural Network.