From electrons to phase diagrams with machine learning potentials using pyiron based automated workflows - Research overview

Aug 4, 2026

A user-friendly pyiron-based framework is presented to automate the full machine learning potential (MLP) development cycle: generating systematic DFT databases, fitting DFT data to empirical or neural and basis-expansion potentials, and validating the resulting models largely automatically. The workflow is demonstrated across three distinct potential classes, including EAM, HDNNP, and ACE. As a validation and application case, the framework computes the Al-Li composition–temperature phase diagram, highlighting transferability and safe deployment through comprehensive physical and dynamical checks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

AccessFree to download

Overview of the MLP development cycle

Workflow stages: DFT databases, fitting, validation

Potential classes covered: EAM, HDNNP, ACE

Advanced application: Al-Li phase diagram

npj | computational materials Article | Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences | | | | --- | --- | --- | | [https://doi.org/10.1038/s41524-024-01441-0](https://doi.org/10.1038/s41524-024-01441-0) | | | | From electrons to phase diagrams with machine learning potentials using pyiron based automated workﬂows
Check for updates | | | | Sarath Menon 1 , Yury Lysogorskiy2, Alexander L. M. Knoll 3,4, Niklas Leimeroth 5, Marvin Poul 6, Minaam Qamar2, Jan Janssen6, Matous Mrovec 2, Jochen Rohrer 5, KarstenAlbe 5, Jörg Behler 3,4,

Ralf Drautz 2 & Jörg Neugebauer 6 | | | | We present a comprehensive and user-friendly framework built upon the pyiron integrated development environment (IDE), enabling researchers to perform the entire Machine Learning Potential (MLP) development cycle consisting of (i) creating systematic DFT databases,(ii) ﬁtting the Density Functional Theory (DFT) data to empirical potentials or MLPs, and (iii) validating the potentialsin a largely automatic approach. The power and performance of this framework are demonstrated for three conceptually very different classes of interatomic potentials: an empirical potential (embedded atom method-EAM), neural networks (high-dimensional neural network potentials-HDNNP) and expansions in basis sets (atomic cluster expansion-ACE). As an advanced example for validation and application, we show the computation of a binary composition-temperature phase diagram forAl-Li, a technologically important lightweight alloy system with applications in the aerospace industry. | | | | The advent of machine learning interatomic potentials (MLPs) is revolutionising the ﬁeld of computational materials science, enabling simulations of large systems and complex material properties with ab initio accuracy1–5. However, the development of these data-driven interatomic potentials is a computationally intensive task that needs automated and reliable workﬂows.
The life cycle of MLP development can be broadly divided into the following tasks: (i) generating a database containing reference data, (ii)ﬁtting the model parameters to the reference data, and (iii) validating the resulting parametrization fora speciﬁed range ofproperties. Furthermore, it is often necessary to provide a feedback loop between the tasks via an active learning approach to ascertain transferability6,7.
The initial task of setting up the reference database usually requires to perform many thousands of density functional theory (DFT) calculations for a broad range of atomic environments that span the conﬁguration space of interest as completely as possible. Such computations can nowadays be facilitated using either general workﬂow frameworks8–11 or tools designed speciﬁcally for a particular MLP class12–17. Nevertheless, there is still lack of standardized workﬂow setups, computational metaparameters and structural databases. Therefore, each research group relies mostly on their own | expertise and experience. This may not only lead to inconsistencies in the generated data (for instance, due to variations in DFT settings, such as the exchange-correlation functional, Brillouin zone sampling or plane wave cutoff)18, 19, but it can also strongly limit exchange of data from different sources and their collection into greater databases.
The situation remains similar when it comes to the second stage of MLP development, namely, ﬁtting the model parameters. Many optimization algorithms and software tools are tailored to a particular class of potentials and are not easily transferable. Thus, a researcher must not only identify an appropriate type of potential that is suited for the system of interest, but often needs to learn a variety of speciﬁc software tools each of which uses its own terminology.
The ﬁnal task in the development cycle is a thorough validation of the ﬁtted parametrization. It should be stressed that simple correlations be

FAQs

What are the main stages of the machine learning potential (MLP) workflow presented?

The framework automates (i) creating systematic DFT reference databases, (ii) fitting DFT data to empirical potentials or MLPs, and (iii) validating the fitted models for specified property ranges.

Which types of interatomic potentials are demonstrated in the framework?

Three conceptually different classes are used: an empirical potential (embedded atom method, EAM), neural networks (high-dimensional neural network potentials, HDNNP), and basis-set expansions (atomic cluster expansion, ACE).

How is the framework validated in an advanced application example?

It computes a binary Al-Li composition–temperature phase diagram, serving as a validation and practical application that depends on trustworthy model behavior beyond basic correlations.

Aug 4, 2026

Upload by Seraphina