AVS 71 Session AIML-ThP: AI/ML for Scientific Discovery Poster Session

Over the past year there has been an increasing interest in leveraging foundation and large language models to design AI agents that can interact withexperiments to solve materials science and synthesis problems. One of the challenges of this approach is that testing the performance of these agents require access to automated labs. In contrast to benchmarks testing abilities such as knowledge, math skills, or reasoning, there is a lack of benchmarks that can help both design and evaluate agents without the access to dedicated experimental facilities.

In this work, we introduce Semibench, a benchmark to evaluate AI agents' ability to operate and solve synthesis challenges in the context of semiconductor processing. This benchmark introduces two core ideas: first, it introduces virtual tools that simulate the output of real life experiments. This allows us to test an agent's ability to solve a wide range of challenges involving different tool configurations, amount and nature of information that is accessible, and process complexity. Second, it focuses on the concept of microtasks, challenges designed to have a unique solution. This allows us to define quantitative performance metrics for the agent based on how far the proposed solution is to the ground truth. For Semibench, we have focused on three different techniques that are commonly used in the context of microelectronics: atomic layer deposition, sputtering, and reactive ion etching. For each challenge in the benchmark, agents are exposed to a collection of virtual tools and asked to solve specific questions by providing a sequence of synthesis steps. These steps involve selecting the right configurations for each of the tools, such as the precursor channels in the case of ALD, or the targets and power for sputtering, or the etching recipe for RIE.

Machine learning interatomic potentials (MLIPs) are powerful tools for accelerating atomistic simulations, but their reliability depends critically on training set construction. A common strategy is to build universal MLIPs from open-source databases, offering transferability but often sacrificing accuracy particularly in surfaces and interfaces with highly diverse chemical environments. These databases are dominated by equilibrium structures, leaving reaction pathways undersampled, forcing potentials to extrapolate in chemically critical regions—a limitation in catalysis, surface chemistry, and defect dynamics where reactive events dominate. Specialized MLIPs built via active learning can achieve higher accuracy but typically rely on molecular dynamics (MD) and committee models to sample configuration space. Because rare reactions are unlikely to appear within accessible timescales, discovery is left to chance, often requiring long simulations or many iterations. As a result, such MLIPs may fit training data well but fail to capture the most chemically relevant regions.

To address this limitation, we introduce DIAL (Domain-Informed Active Learning), a chemically targeted strategy that augments conventional active learning. Rather than relying solely on MD and uncertainty-driven sampling, DIAL incorporates both established reaction pathways and those identified via nudged elastic band (NEB) calculations. Training datasets are enriched with configurations along these pathways, particularly near transition states. By integrating data-driven active learning with domain knowledge of chemical processes, this approach ensures that the potential is trained on the chemically important regions of configuration space.

Using DIAL, we developed a specialized MLIP for diamond surfaces and interface reactions. The potential enabled large-scale molecular dynamics simulations that reproduced graphitized and oxidized surface morphologies and reaction products, while providing new insights relevant to diamond-based electronics and quantum technologies. In particular, the model captured thermal degradation mechanisms and suppression, facet-dependent oxidative etching, and suggested strategies for controlling surface termination to preserve quantum-relevant color centers. Although demonstrated on diamond, the DIAL framework is general and applicable to other reactive materials systems, including catalysis and battery interfaces.

These results demonstrate how DIAL bridges data-driven methods with domain expertise, highlighting the value of collaboration between materials scientists and AI specialists in advancing the next generation of materials discovery.

Advanced plasma etch and deposition process development in semiconductor R&D requires the simultaneous optimization of dozens of interdependent parameters against stringent nanometer-scale metrics. Today, process engineers face a fragmented workflow characterized by siloed data, inefficient experimentation, manual analysis, and limited integration between process settings and physical outcomes. These bottlenecks slow discovery, drive up costs, and hinder knowledge transfer across teams.

We present Elucida Labs, an AI-native operating system designed to transform semiconductor R&D environments across both etch and deposition. Our system enables process teams to reduce experimental burden, shorten learning curves, and converge to target specifications faster. By embedding AI-driven intelligence directly into R&D workflows, Elucida Labs demonstrates how AI can amplify human expertise, accelerate innovation, and reshape the economics of semiconductor process development.