Report on the outcomes of a Short-Term Scientific Mission

Action number: CA21101

Applicant name: Jon Eunan Quinlivan Domínguez

Details of the STSM

Title: Evaluation of machine-learning potentials for accelerating grand canonical global optimization algorithms

Start and end date: 01/10/2023 to 28/10/2023

Description of the work carried out during the STSM

During the Short-Term Scientific Mission, the grantee has developed a Grand Canonical Global Optimization module for the AGOX library, under development by Prof. Bjørk Hammers’s group (host). To develop this module, the grantee has received formation in different Machine-Learning techniques developed in the host group, as well as several applications carried out by other members of the group, in the form of weekly seminars.

Additionally, the applicant has been directly supervised by Mads-peter V. Christiansen, a member of the group developing the local Gaussian Process Regression (local GPR). The local GPR model is trained on local descriptions of the atomic structure, obtained with the Smooth Overlap of Atomic Positions (SOAP) descriptor. Combining this with ab initio Thermodynamics, the model can evaluate the relative stability of structures with different stoichiometries.

Initially, the work of the grantee consisted of adapting the pre-existing AGOX package to be able to work simultaneously with more than one stoichiometry. After this was completed, the applicant included an ab initio Thermodynamics module, which allows the program to evaluate the Gibbs free energy of formation of the considered stoichiometries, with common reference energies.

Following this, the applicant implemented new functionality, regarding the selection and generation of candidate structures based on their Gibbs Free energy of formation, into the AGOX package. The contributions are listed below:

  • Gibbs free energy of formation acquisitor.
  • Gibbs free energy of formation sampler.
  • Atom addition/removal operator.
  • Crossover operator.

The grantee also performed tests to assess the performance of the developed module, for which multiple runs of the Grand Canonical Global Optimization module of the AGOX package with a simple problem system (C2HX X=2-6) were carried out using high-performance computing resources. The performance of the code was measured as a cumulative success curve detailing the percentage of runs that found the global minimum of the system as a function of the number of evaluated structures.

Finally, the grantee presented past and current projects related to global optimization of atomic structures in which he is participating, the current state of the developed code, and future challenges and approaches to the problem at hand.

Description of the STSM main achievements and planned follow-up activities

The scientific objectives of the “COST Action 21101Confined molecular systems: From a new generation of materials to the stars” is to provide computational and experimental building blocks for a fundamental understanding of confined molecular systems: An important first step for the modelling of such materials is to find their stable structure and stoichiometries under reaction conditions. In this context, developing ML potentials and implementing new global optimization algorithms able to predict the state of working nanostructured materials under reaction conditions will significantly advance the objectives of the action.

This STSM has contributed to the 1st Grant Period Goals of:

– WG1: Testing of theoretical tools aimed to calculate interaction potentials between molecular systems and complex environments through composite DFT/ab inito approaches and ML-based representations. Explanation: This corresponds to the implementation and testing of the Grand Canonical Global Optimization (GCGO) module of the AGOX package. A Machine-Learning assisted global optimization algorithm.

– WG2: Testing and training on computational methods addressing the dynamics of molecular systems in molecular cages (MOFs, COFs), surfaces (metal/metal-oxide), interfaces, and under strong electromagnetic static or optical fields. Explanation: The development and application of The AGOX GCGO approach should contribute to identifying stable stoichiometries and structures (minima) that will be visited during the dynamics of relevant systems, and which can be used as starting points for molecular dynamics simulations of relevant systems.

The collaboration between the host group and the applicant group is still ongoing, as more development can still be carried out for the GCGO module of AGOX. These further developments will consist on more advanced candidate selection and sampling techniques, which should further improve the performance of the code.

– WG3: Selection of sample materials to address the synthesis and characterizations of single/few-atoms metal and metal-oxide clusters and their opto-electronic and functional properties, focusing on the impact of the cluster/support interaction. Explanation: We typically target freestanding and supported metal or oxide clusters, which allows us to identify relevant structures and oxidation states and the effect of support interactions on them (a necessary first step before characterizing opto-electronic properties). In relation to this, once the development of the AGOX package is completed, it will be employed to characterize systems of interest such as heterogenous catalysts with varying amounts of H and O atoms at the density functional level of theory, systems in which the applicant group has previous experience.

The implementation of the novel code will also spur collaborations with partners of the action needing to characterize the structure and chemical state of their target systems under reaction conditions. As such, I plan on attending meetings to present this work, hoping to establish collaborations with groups interested in using the library.

The work described above should lead to publications regarding both the developed methodology and the newly characterized systems.

Share this article, choose your platform!