Tejs Vegge

AI-orchestrated materials discovery and closed-loop synthesis of nanoparticles and electrocatalysts

Establishing a distributed infrastructure for autonomous materials discovery and synthesis plays a critical role in accelerating the development of advanced energy materials in areas like sustainable batteries and electrocatalysts for the green transition. A central element in this process is the development of a closed-loop infrastructure or materials acceleration platform (MAP) [1,2], where different nodes, methods, and even geographically distributed laboratory equipment can work jointly using autonomous workflows [3] to co-optimize materials and device-level properties. Here, we show an example using the Fast INtention-Agnostic LEarning Server (FINALES) framework to orchestrate a two-pronged optimization task, where both optimization tasks vary the composition of a battery electrolyte composed of ethylene carbonate (EC), ethyl methyl carbonate (EMC), and lithium hexafluorophosphate (LiPF6). One targets the optimization of ionic conductivity, while the other aims to maximize the end-of-life (EOL) of coin cells [4].

Another key challenge is the controlled synthesis of materials with a specific atomic structure. This underpins many technological advances, yet remains reliant on iterative, trial-and-error approaches. Nanoparticles (NPs), whose atomic arrangement dictates their emergent properties, are particularly challenging to synthesize due to numerous tunable parameters. Here, we introduce an autonomous approach explicitly targeting the synthesis of atomic-scale structures, which autonomously designs synthesis protocols by matching real-time experimental total scattering (TS) and pair distribution function (PDF) data to simulated target patterns, without requiring prior synthesis knowledge. We demonstrate this capability at a synchrotron, successfully synthesizing two structurally distinct gold NPs: 5 nm decahedral and 10 nm face-centred cubic structures. Ultimately, specifying a simulated target scattering pattern, thus representing a bespoke atomic structure, and obtaining both the synthesized material and its reproducible synthesis protocol on demand may revolutionize materials design. Hence, ScatterLab provides a generalizable blueprint for autonomous, atomic structure-targeted synthesis across diverse systems and applications [5]. We also show how the application of a fine-tuned version of the MACE-MP-0 foundation model [6] enables us to simulate the NP-synthesis conditions and guide the synthesis by reaching spatio-temporal scales that are outside the realm of traditional ab initio molecular dynamics simulations.

As a third example, we present FastCat - an AI-orchestrated self-driving closed-loop materials discovery system for the autonomous discovery of platinum group metal-free multi-metal catalyst for alkaline OER (Oxygen Evolution Reaction) [7]. With FastCat, we have synthesized, characterized, and tested more than 500 Ni-based multielement layered double hydroxide (LDH) catalysts in one of the most extensive AI-orchestrated catalyst discovery campaigns to date. Our metaheuristic Bayesian Optimization identified known high-performance compositions and several novel multielement Ni-Fe-Cr-Co alloys with unprecedented overpotentials at higher current densities.

Finally, we discuss how recent advancements in ML have demonstrated that simple material representations like chemical formulas without any structural information can sometimes achieve competitive property prediction performance in common tasks. Our physics-based intuition would suggest that such representations are “incomplete,” which indicates a gap in our understanding. A tomographic interpretation of structure-property relations is used to bridge that gap by defining what a material representation, material properties, the material, and the relationships between these are [8]. We apply concepts from information theory to verify this framework by performing an exhaustive comparison of property-augmented representations on a range of materials’ property prediction objectives. Thus, as scientists, we might not know a priori which experiments or simulations will ultimately provide the most valuable information to capture the “ghost of the material,” i.e., what is the fastest path to obtain a complex material’s property.

1. Stier et al., Adv. Mater, 2024, https://doi.org/10.1002/adma.202407791

2. Canty et al., Nature Comms, 2025, https://doi.org/10.1038/s41467-025-59231-1

3. Sjølin et al., Digital Discovery, 2024, https://doi.org/10.1039/D4DD00134F

4. Vogler et al., Adv. Energy Mater, 2024, https://doi.org/10.1002/aenm.202403263

5. Anker et al., arXiv, 2025, https://doi.org/10.48550/arXiv.2505.13571

6. Batatia et al., arXiv, 2024, https://doi.org/10.48550/arXiv.2401.00096.
7. Fisker-Bødker, Moralles, Chang, Vegge, ChemRxiv, 2025, https://doi.org/10.26434/chemrxiv-2025-w8bkg.
8. Ortega-Ochoa, Aspuru-Guzik, Vegge, Buonassisi, arXiv, 2025, https://doi.org/10.48550/arXiv.2501.18163