What are Machine-Learned Force Fields?

Machine-learned force fields (ML-FFs) aim to address the system-size limitations of accurate ab initio methods by learning the energies and interactions in atomic-scale systems directly from, for example, density functional theory (DFT) calculations. This is in contrast to conventional force fields (FF), which try to parameterize a fixed analytical approximation of the energy landscape. When trained sufficiently, ML-FFs can predict these interactions for much larger systems and therefore combine the accuracy of ab initio methods for molecular dynamics (MD) simulations with the computational efficiency of conventional force fields. The result is a modern and data-driven approach to accelerating atomic-scale simulations of material properties for virtually any kind of material.

How do Machine-Learned Force Fields work?

As opposed to conventional FFs, ML-FFs are based on mathematical constructions with the very little inherent concept of physics. Therefore, training the ML-FF on relevant high-accuracy DFT data (energies, forces, stress) is key to obtaining a reliable ML-FF for particular systems and applications. During training and simulation, the atomic environments in a configuration are converted into a set of generic descriptors (features), which are fed into the actual machine learning algorithm (e.g. linear regression or neural network) to predict the energies of the atomic configuration. The ML-FF is trained by fitting the parameters in the ML model to minimize the difference between predicted and ab initio energies, forces, and stress in the training data. When the training is completed, the ML-FF model can be used for atomic-scale simulations in the same way as any other conventional FF.

What are the benefits of Machine-Learned Force Fields?

 

A long-standing challenge in atomic-scale MD simulations is that reliable simulation models are either too expensive (ab initio) to reach the simulation time- and length scales that would allow one to capture the complexity of realistic systems, or limited accuracy for relevant systems (conventional FF). The ML-FF approach offers a solution to this challenge and enables, for example, simulation of dynamical atomic-scale processes; these require high accuracy but take place on longer time scales, e.g. diffusion, crystallization, or deposition, such that the user can simulate such processes not only on model systems with significantly reduced complexity but on the more realistic complex systems without loss of accuracy. Moreover, ML-FFs also enable the user to rapidly generate and optimize a variety of different realistic input structures that can be used in, for example, ab initio calculations. This reduces the number of computationally expensive calculations significantly and allows for a larger throughput of different structural parameters, among other benefits.

 

 

About Synopsys Machine-Learned Force Fields

How do Machine-Learned Force Fields fit in the Synopsys product portfolio?

The QuantumATK atomistic modeling platform from Synopsys is one of the most modern and versatile materials simulation codes on the market. It can be used to compute electronic, thermal, mechanical, optical, magnetic, ferroelectric, thermoelectric, and many other properties of complex molecules, amorphous and crystalline materials, interfaces, and even atomistic devices.

QuantumATK implements the moment-tensor-potential (MTP) class of ML-FFs, which is a modern, state-of-the-art algorithm that provides an excellent balance between accuracy and efficiency. One major advantage of QuantumATK is that several materials simulation frameworks are available in one unified platform, including DFT, tight-binding methods, and force fields. Advanced ML-FF training workflows, which require both DFT and FF simulations, are therefore created and executed within a single software platform.

QuantumATK also ships with custom workflows explicitly designed to minimize the number of DFT calculations needed to generate a reliable ML-FF, and automates most parts of the training workflow. The graphic below illustrates the main steps in the ML-FF training workflow for the MTP in QuantumATK:

  1. Generate training configurations using one of the efficient automatic protocols in QuantumATK
  2. Compute DFT training data for these configurations
  3. Train the ML model to the generated training data
  4. Use the final MTP model in simulations like any other FF

Going beyond Machine-Learned Force Fields

As the semiconductor industry continues advancing towards more advanced processing nodes, atomic-scale details of device interfaces and nanoscale thermal transport properties become increasingly important to device and processing design. Several Synopsys TCAD tools are used to address these challenges, and the ML-FF methodology is key to connecting atomistic simulations to many of these tools, for example by providing atomic-scale insight into the detailed composition of high-K metal gate stacks for advanced logic or for computing thermal properties of novel materials and interfaces.

Putting Machine-Learned Force Fields into Practice

QuantumATK Case Studies

QuantumATK’s pre-trained moment-tensor-potentials in the Multilayer Builder can be used to generate and optimize close-to-reality atomic-scale configurations for the interfaces in High-K Metal Gate stacks. The resulting configurations can be used to study electronic or structural properties, such as to investigate the effects of layer thickness, defects, or dopants.

Learn more about QuantumATK products

Interested in applying QuantumATK software to your research? Test our software or contact us at quantumatk@synopsys.com to get more information on QuantumATK platform for atomic-scale modeling.