Posters at this year's conference

We present a fully automated system for catalyst synthesis and analysis based on a conveyor belt platform. It combines microfluidic channel, thermal treatment, and in-line SECCM/XRD modules, enables to find adequate catalyst used for enhancing OER performance.

We propose a method for predicting the surface profile shape in coating processes using statistical shape analysis and Fréchet regression. This method can be used as a tool for data-driven optimization to achieve optimal coating layer shapes.a

We present an optimization method for designing microfluidic concentration gradient generators. The complex internal structure of the microfluidic device makes computation difficult. We propose an electric circuit analogy-based channel network computation with structure optimization to overcome this problem.

We perform loss landscape analysis on neural network models in the context of active learning. Our results provide both visual and quantitative evidence of progressive improvements in model stability and generalizability across active learning iterations. Furthermore, we explore the potential of loss landscape analysis to inform training set pruning and guide the development of novel acquisition policies.

This project focuses on developing an intuitive teleoperation system to collect high-quality human demonstration data for robot learning. By enabling force feedback and retargeting control, the system aims to facilitate more natural human-robot interaction for training robots in contact rich lab tasks.

A self-driving workflow guided by Bayesian optimization accelerated the discovery of inherently recyclable polymer resins by integrating dynamic covalent monomers such as vinylogous urethanes with comonomers to achieve targeted mechanical properties. This automated approach enabled rapid exploration and synthesis of polymer compositions, advancing the development of more sustainable and reprocessable materials.

WATCHDOG is an end-to-end framework for fault detection and monitoring in robotics-based self-driving lab workflows, powered by real-time digital twin simulations. We deploy WATCHDOG in an automated electrochemistry workflow with multiple robots and custom labware, demonstrating its ability to prevent errors and improve robustness to experimental variability.

We present a freely available Excel/Google sheets no-code plug-in to run Bayesian Optimization. Designed for both manual as well as fully automated self-driving laboratories, it leverages the battle-tested user-interface mechanism of Excel/Google to solve a core adoption challenge of Bayesian Optimization: reliable data storage and handling.

LISA is a robotic lab assistant that helps scientists organize and prepare their experiments, automatically tracks inventory in real-time, and reduces repetitive tasks- allowing researchers to focus more on the science itself.

We're building a robotic workflow that can automate the testing and discovery of sustainable degradation for next-generation electronic materials. This study compares how well the robot performs against human scientists in this process.

This study presents a data-driven and generative approach to efficiently discover high-anisotropy magnetic materials by constructing a reliable MCAE database based on stable ferromagnets from the Materials Project. A two-stage machine learning pipeline—classification followed by regression with uncertainty quantification—enables effective screening and supports future active learning. The search space is further expanded via a conditional generative model that proposes promising unseen structures beyond existing databases.

We propose an LLM-based framework that redesigns synthetically infeasible inorganic crystal structures into experimentally realizable ones using a text-encoded CIF representation. The model successfully converts over 3,395 structures, with a 34% match rate to known experimental structures. This highlights the potential of LLMs in bridging computational predictions with real-world synthesis for AI-driven materials discovery.

We show how an old AI method for making money on slot machines can cut the supercomputing costs of designing clean energy materials.

We present a flexible data-driven framework to guide high-throughput late-stage functionalization of drug compounds. This pipeline generates diverse therapeutic candidates in silico with promising properties and synthesizability via user-defined late-stage functionalization reactions with predicted conditions.

Proton exchange membrane water electrolysis (PEMWE) is a key pathway for green hydrogen production, but its efficiency is limited by the sluggish and costly oxygen evolution reaction (OER) under acidic conditions. To accelerate the discovery of acid-stable OER catalysts, we developed an AI-driven workflow focused on metal/metal oxide (MMO) interfaces that stabilize active sites and improve performance. By integrating DFT, automated synthesis, and machine learning, we investigate structure–function relationships and target both conventional and emerging OER mechanisms to design catalysts that are active, durable, and cost-effective.

We’re developing a smart digital tool that helps pharmaceutical companies design tablet formulations more efficiently by predicting how ingredients will compact together to cut down on waste, cost, and the amount of drug needed. This method uses a Model-Based Design of Experiments and Optimization to help formulation scientist make decisions with fewer experiments, speeding up development and supporting more sustainable manufacturing.