Matteo Manica

Heterogeneous catalyst design is primarily expert-driven and difficult to keep track amidst rapidly growing literature, motivating the use of large language models (LLMs) to learn from existing knowledge. While generalist LLMs show some capability in generating synthesis procedures, they struggle on encountering complex catalytic systems like single-atom catalysts (SAC). Here, we investigate fine-tuning of LLMs to improve the generation of SAC synthesis procedures with chemical relevance. We curate a dataset of 2,964 SAC publications and train the Granite-based LLM to recommend protocols based on user-defined prompts including metal-support combinations, synthesis method, and target reactions. We demonstrate the model’s practicality through a user interface allowing researchers to query procedures tailored to their design conditions. Human-in-the-loop validation of 150 model-generated SAC protocols confirmed coherent, context-aware procedures with synthesis sequence steps that align with experimental logic, along with limitations. We underscore LLM-generated protocols must be critically reviewed and used as assistive tools, and highlight fine-tuned LLMs for AI-assisted synthesis planning driving innovation in catalysis research.

The concept of self-driving chemical laboratories builds on top of two technologies: robust artificial intelligence models that translate synthesis designs into executable instructions, potentially incorporating instrument feedback, and robotic workstations capable of executing those procedures seamlessly and without human intervention. Here we present RoboRXN, a cloud-native framework that unifies transformer-based retrosynthesis and chemical action extraction or prediction models with a commercially available synthesis robot and in-line high-resolution LC–MS. Across more than 3000 actions autonomously executed, we analyzed in details 33 reactions distributed among 10 single-step reactions and 9 multi-step campaigns. Eightythree percent of single-steps and all two-step routes succeeded without parameter adjustment; a branched four-step synthesis completed after one on-deck modification. Regardless of the overall success, we systematically encountered three recurring bottlenecks: hardware rigidity, execution-level data gaps, and slow physical feedback, which are seldom reported in the literature, despite being common to many commercial platform design. We delineate how each failure surfaced, quantify its impact on yield or throughput, and outline engineering and data-centric remedies. By publishing both the successes and the blind spots, we hope to redirect community effort from one-off demonstrations toward robust, transparent, and widely adoptable AI-driven laboratories.

Bridging AI and self-driving laboratories, we introduce the first fully-automated, closed-loop molecular discovery cycle, exemplified by the identification of novel JAK inhibitors. With minimal human intervention, we combined AI-driven molecular design and retrosynthesis with IBM’s synthesis automation system RoboRXN and Arctoris’ Ulysses platform for automated in-vitro screening. We performed two Design-Make-Test-Analyze (DMTA) cycles with a total of 36 synthesized compounds. In the first round, the generative model identified structural analogs of established JAK inhibitors despite being blinded from known JAK inhibitors. The second iteration yielded molecules with significantly improved pIC50 and ligand efficiency (p < 0.001) compared to the first round, underscoring the effectiveness of the closed-loop workflow.

Combining AI and self-driving laboratories unlocks closed-loop molecular discovery by automation across all stages of the Design-Make-Test-Analyze (DMTA) cycle.

Transformers have proven successful in a range of sequence modelling tasks. However, these models have significant limitations: they are inherently data-greedy, and suffer from the risk of training data leakage. These limitations prevent their broad application in various domains. While the advent of foundation models (FMs) addresses the data-greedy nature of Transformers, the risk of exposing training data remains; it has been demonstrated that excerpts of the training data can be obtained by prompt engineering on an FM. To simultaneously address these limitations, we propose unified lookup tables (ULTs), a data preprocessing step that enables building and fine-tuning FMs on encoded data. ULTs enable the reuse of a trained model on new datasets without exposing any unencoded training data. The method leverages data compression methods as efficient modality tokenizers, and a common representation vocabulary to facilitate fine-tuning on encoded data. We theoretically support our claims through numerical estimations of the likelihood of reverse engineering the data encoding and practically through empirical evaluation on domains that can benefit from ULTs. Specifically, we evaluate the impact of using ULTs as a preprocessing step before training both decoder-only and encoder–decoder language models on text, images, and molecules. We demonstrate that the encoding step does not negatively affect model training and leads to an average relative increase of ∼16% on a collection of text metrics, while producing close to competitive results on image classification and chemical reaction prediction tasks. Code to reproduce the experiments is available at: https://github.com/IBM/unified-lookup-tables .

Large language models, commonly known as LLMs, are showing promise in tacking some of the most complex tasks in AI. In this perspective, we review the wider field of foundation models—of which LLMs are a component—and their application to the field of materials discovery. In addition to the current state of the art—including applications to property prediction, synthesis planning and molecular generation—we also take a look to the future, and posit how new methods of data capture, and indeed modalities of data, will influence the direction of this emerging field.

Vision-language foundation models augmented with domain-specific data can generate documentation of scientific experiments in real time based on egocentric videos of laboratory actions.

Language model assistants have transformed how researchers interact with computational tools, offering unprecedented capabilities in understanding and generating complex scientific queries. We introduce a language model assistant for biocatalysis (LM-ABC), a computational tool designed to streamline workflows in enzyme engineering research. LM-ABC integrates a large language model with domain-specific modules to facilitate biocatalysis research through natural language inputs. Its architecture employs the Reasoning and Acting (ReACT) framework for dynamic tool selection and chaining, enabling functionalities like binding site extraction and molecular dynamics simulations. LM-ABC can interpret and process user queries in the form of natural language, and interface with existing computational resources to generate relevant results for enzyme engineering. Additionally, LM-ABC is available via both command-line and web-based interfaces, which lowers the barriers for its usage and integration in various disciplines. Provided as open-source software, the LM-ABC contributes to the application of language models in computational biology, potentially accelerating enzyme engineering research processes.

Combining chemical sensor arrays with machine learning enables designing intelligent systems to perform complex sensing tasks and unveil properties that are not directly accessible through conventional analytical chemistry. However, personalized and portable sensor systems are typically unsuitable for the generation of extensive data sets, thereby limiting the ability to train large models in the chemical sensing realm. Foundation models have demonstrated unprecedented zero‐shot learning capabilities on various data structures and modalities, in particular for language and vision. Transfer learning from such models is explored by providing a framework to create effective data representations for chemical sensors and ultimately describe a novel, generalizable approach for AI‐assisted chemical sensing. The translation of signals produced by remarkably simple and portable multi‐sensor systems into visual fingerprints of liquid samples under test is demonstrated, and it is illustrated that how a pipeline incorporating pretrained vision models yields average classification accuracy in four unrelated chemical sensing tasks with limited domain‐specific training measurements. This approach matches or outperforms expert‐curated sensor signal features, thereby providing a generalization of data processing for ultimate ease‐of‐use and broad applicability to enable interpretation of multi‐signal outputs for generic sensing applications.