Chemists will soon have access to thousands of open datasets designed specifically to train artificial intelligence models, generated by robotic platforms and large language models at Imperial College London. The problem is straightforward: AI models in chemistry are only as good as the data they learn from, and much of the existing data is sparse, poorly structured, or locked behind paywalls. This project fills that gap by producing three types of high-quality datasets. First, experimental data from closed-loop robotic optimisation platforms, including logs from AI systems that explain their own reasoning. Second, a large, uniformly sampled dataset for chemical reaction optimisation, capturing multi-objective and multi-fidelity data. Third, a multimodal dataset combining NMR, mass spectrometry, and FTIR data for roughly 2,000 molecules, plus a synthetic dataset of about 700,000 molecular structures. If successful, these resources will accelerate the adoption of AI tools across the chemical sciences. Researchers could use them to predict molecular structures from experimental spectra, optimise chemical reactions faster, or develop generative models for drug discovery. The datasets are openly accessible, so any lab—academic or industrial—can use them without barriers. This is fundamental infrastructure work: it does not produce a new drug or material directly, but it removes a key bottleneck that has kept AI from transforming chemistry the way it has transformed image recognition or natural language processing.
View original technical description
This project will deliver high quality, diverse, and openly accessible datasets that address bottlenecks in the application of artificial intelligence (AI) to chemistry. These datasets will support the development, benchmarking, and validation of AI models and methodologies, particularly in areas where robust, well structured, and representative data is currently lacking. This aligns directly with a core mission of the AIchemy Hub, to unlock the potential of AI in the chemical sciences by making foundational data resources available to the broader community. Our approach focuses on three interrelated and high impact areas. First, we will produce high dimensional experimental datasets from closed-loop robotic optimisation platforms under three distinct regimes: i) ‘vanilla’ Bayesian optimisation, (ii) Human-in-the-Loop approaches (HypBO), and (c) LLM reasoning (BORA). These datasets will be complemented by rich textual logs generated by the LLMs, enabling future studies of interpretability, decision rationale, and human-AI interaction. Secondly, we will create a large, uniformly sampled dataset for chemical reaction optimisation. This will capture multi-objective and multi-fidelity data, essential for developing and evaluating advanced optimisation algorithms. By utilising the ROAR and ATLAS facilities at Imperial College London we will ensure datasets are both comprehensive and reproducible, enabling its use as a robust benchmark for the community. Thirdly, we will assemble a multimodal dataset designed to bridge the gap between theoretical simulations and experimental data in molecular identification. This dataset will include NMR, mass spectrometry, and FTIR data for ~2,000 molecules, supplemented with a large synthetic dataset of ~700,000 molecular structures. Such data are urgently needed for training generative and predictive models capable of interpreting experimental spectra and deriving molecular structures in complex environments. The outputs of this project will form the foundation for a range of activities including datathons, hackathons and cross disciplinary tool development. By filling known data gaps with high value, openly available resources, this project will catalyse the adoption of AI tools in chemistry and foster new collaborations and promote interdisciplinarity across the physical sciences.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know