Identification number: LZP-2024/1-0406

Type: Latvian Council of Science Fundamental and Applied Research project

Duration: 01.01.2025. - 31.12.2027.

Project Leader: Dr. Sergejs Piskunovs, Institute of Solid State Physics University of Latvia (ISSP UL)

Total funding: 300 000 EUR


 

About the project

The project started on 1 January 2025 at the Institute of Solid State Physics, University of Latvia (ISSP UL). It aims to accelerate the discovery of novel heterojunction photocatalysts for efficient green hydrogen production by combining generative machine learning (ML) with density functional theory (DFT/TDDFT) calculations. Two-dimensional materials from the Computational 2D Materials Database (C2DB) serve both as the main target materials and as training data for the ML models, while targeted synthesis and photocatalytic testing close the design loop. The project pursues six research objectives: (i) a DFT-supported generative ML framework for the rational design of heterojunction photocatalysts; (ii) identification of the key structural and electronic descriptors that govern photocatalytic activity; (iii) an openly accessible (FAIR) database of photocatalysts and their properties; (iv) simulation of interfacial charge-transfer dynamics by non-adiabatic molecular dynamics; (v) optimisation of heterojunction structures and compositions with intelligent algorithms; and (vi) experimental validation of the most promising candidates in cooperation with the University of Oulu (Finland) and the National Taipei University of Technology (TaipeiTECH, Taiwan). The results are expected to speed up the development of efficient and cost-effective photocatalysts for green hydrogen and to support the transition to sustainable energy.


 

Progress after the mid-term report

The project is proceeding according to the work plan. The DFT-supported generative ML framework has been completed and validated: graph neural networks predict band gaps, band-edge positions and formation energies directly from the crystal structure, generative models propose new candidate heterojunctions, and an active-learning loop between DFT and ML continuously improves the accuracy of the predictions. Ensemble ML models showed that band-edge alignment, band-gap width, dopant concentration, local structural distortion and nanostructure morphology have the greatest influence on charge separation and visible-light absorption. For MXene-based materials, an ML workflow was developed to predict the CO₂ adsorption energy, a key quantity for photocatalytic CO₂ conversion. A structured database of heterojunction photocatalysts has been built; it includes a screening set of about 4000 MXene configurations, roughly 1000 of which were calculated explicitly. Hybrid DFT calculations revealed how oxygen vacancies in ZrO₂ create defect states that change its electronic structure and optical absorption. A Bayesian-optimisation and genetic-algorithm loop predicted the most favourable composition range for Mn-doped BiVO₄ photoanodes, and the first experiments at TaipeiTECH confirmed the prediction: Mn doping narrows the band gap from 2.44 eV (pristine BiVO₄) to 2.11 eV (x = 0.5) and improves the band alignment for water oxidation.

To date, two joint open-access review articles have been published (Molecules, 2026; Catalysts, 2026), and three posters have been presented at international conferences (E-MRS 2025 Spring Meeting, FM&NT 2026 and EURODIM 2026). Cooperation has been established with five research groups in Greece, Finland, Taiwan and Kazakhstan, and access to the Taiwanese supercomputer Forerunner1 (3.75 million CPU core hours) substantially accelerated the large-scale calculations. Building on the project’s results, ISSP UL, together with partners from Taiwan, Lithuania, Poland and Israel, including an industrial partner, has submitted the pre-proposal ENHANCE-2D to the M-ERA.NET Call 2026. A doctoral thesis on the project topic is in preparation. In the second half of the project (2026–2027), the work will focus on charge-transfer simulations at heterojunction interfaces, multi-objective optimisation, further experimental validation of the predicted photocatalysts and opening the database to the research community.