Machine Learning Achieves 95% Accuracy with K-Point Mesh Generation Optimized for Quantum ESPRESSO

Although accurate and efficient materials modeling relies heavily on density functional theory calculations, determining the optimal computational settings remains a major challenge, especially for large-scale studies. Elena Patyukova, Junwen Yin, and Susmita Basak from the Department of Scientific Computing at the Daresbury and Rutherford Appleton Institutes are now tackling this problem with a new machine learning approach. The team has developed a system that automatically generates input files for Quantum Espresso calculations, highly optimizing the k-point mesh, a parameter that controls simulation accuracy and computational cost. By training the model on a dataset of over 20,000 materials, we achieve reliable predictions of appropriate k-point settings, ensuring converged results for the majority of compounds and streamlining material discovery workflows. This advancement is expected to accelerate materials modeling by eliminating important bottlenecks in computational efficiency and accuracy.

This is a particularly important issue for high-throughput agent workflows, where additional convergence studies are preferably avoided due to computational cost. Therefore, tools and models that can predict density functional theory (DFT) parameters from basic input information such as structure are needed.

Predicting DFT convergence parameters using machine learning

Scientists have developed a machine learning approach to predict optimal convergence parameters for density functional theory (DFT) calculations, a cornerstone of computational materials science. Choosing appropriate parameters, such as k-point density, often requires time-consuming trial and error. This work aims to automate and optimize this process and build a model that reliably and efficiently predicts these parameters for material simulation. The key innovation is to ensure reliability of predictions using equated quantile regression, a statistical method that provides prediction intervals with guaranteed coverage probabilities.

This allows scientists to assess the reliability of predictions and make informed decisions. The team investigated a variety of machine learning models, including random forests, gradient boosting machines, graph neural networks, and attention-based networks. These models leverage a variety of descriptors to represent materials, including compositional information, crystal structure, and even insights extracted from the materials science literature. The focus extends beyond simple predictive accuracy to robust quantification of uncertainty, which is essential for high-throughput computing and efficient materials screening. By making code and data publicly available, the authors promote reproducibility and collaboration within the scientific community.

Predicting DFT K-point convergence using machine learning

Scientists have developed a machine learning approach to predict optimal parameters for density functional theory (DFT) calculations, with a particular focus on k-point sampling within the Quantum ESPRESSO software package. This addresses an important need in high-throughput materials science, where large-scale convergence studies to determine appropriate parameters are computationally expensive and time-consuming. The researchers generated a comprehensive training dataset consisting of more than 20,000 materials to facilitate accurate model training and validation. Each material was evaluated with an energy convergence threshold of 1 meV/atom. The team evaluated several machine learning models to predict k-point distances, favoring models that could estimate the prediction uncertainty.

Importantly, the model is designed such that for at least 85% of the compounds, the predicted k-point distances fall within the convergence region, minimizing the risk of inaccurate results. Further improvements included predicting specific quantiles and ensuring that high proportions of predictions are not underestimated, thereby balancing accuracy and computational cost. Experiments demonstrate the successful development of a practical web application that automatically generates input files for single-point calculations. This tool streamlines DFT workflows, improves the quality of high-throughput datasets, reduces unnecessary computational time, and contributes to more efficient and sustainable materials discovery. This work significantly advances the automation of DFT parameter selection, enabling more robust, reliable, and high-throughput calculations.

Predicting K-point convergence using machine learning

In this work, we present a machine learning approach to predict appropriate k-point sampling for density functional theory calculations, addressing an important challenge in high-throughput materials science. The team developed a model trained on a dataset of more than 20,000 compounds, allowing them to reliably predict the k-point distance required to achieve energy convergence. By accurately estimating the uncertainty, the model ensures reliable predictions for a large portion of the material (typically 85-95%). The resulting model represents an important step towards automating parameter selection in computational workflows, potentially reducing computational costs and increasing efficiency in materials discovery.

The team made these models publicly available through a web application, facilitating widespread adoption and use within the scientific community. Although this study focuses on single-point computations, the authors acknowledge that extending the approach to other types of computations and properties is an area for future development. They also say that further research could consider integrating these predictive models within new AI-driven computational frameworks. This work is expected to accelerate materials discovery by reducing the computational burden associated with parameter optimization and improving the reliability of high-throughput calculations.