Simon D. Rihm

Artificial intelligence can propose materials, interpret measurements and increasingly coordinate experiments. Yet reliable scientific autonomy requires more than capable models: research systems must preserve what is known, where evidence came from and which actions are justified. We argue that machine-actionable scientific knowledge is essential for trustworthy autonomous research, and that ontologygrounded knowledge graphs are a particularly powerful way of achieving it when scientific meaning must be shared across disciplines, systems and scales. Using The World Avatar (TWA) as a case study, we examine how ontologies, dynamic knowledge graphs and software agents connect models, literature, experiments and laboratory assets, and how ontology-derived tools can constrain large language model (LLM) actions. A reported design-to-synthesis study of a new metal–organic polyhedron demonstrates the integration of previously separate tasks, but not yet an environmental benefit or fully autonomous discovery. A material-passport example further shows how shared semantics can support industrial circularity assessment. We distinguish structural validation from scientific verification, and circularity indicators from demonstrated environmental benefits. The priority is not to create a universal platform but to build evidence-preserving research infrastructure whose scientific and sustainability benefits can be tested.

Metal-organic polyhedra (MOPs) offer a modular route to porous molecular materials, but their discovery is limited by the difficulty of translating digital designs into experimentally accessible compounds. Here, we demonstrate an integrated workflow for predicting and synthesising novel MOPs in The World Avatar. To our knowledge, this is the first reticular material discovery workflow executed and persisted entirely in a shared machine-readable knowledge representation, without human-specified initial reaction conditions. This domain’s small literature corpus hinders data-driven synthesis prediction, yet its high symmetry provides exploitable structural analogies. We therefore encode auditable chemical reasoning as explicit rules rather than learning it statistically. MOP candidates are linked to synthesis knowledge, so procedures for new targets are inferred by analogy to known materials and rendered as human-readable instructions and machine-executable protocols. Candidates are assessed through hierarchical geometry optimisation, electronic-structure evaluation, crystal structure prediction, and simulated powder X-ray diffraction. We evaluate the workflow retrospectively against reported zirconium MOP syntheses and prospectively by synthesising a previously unreported Zr-EDB-MOP. Infrared spectroscopy, high-resolution mass spectrometry, and powder X-ray diffraction support formation of the targeted cage in agreement with the predicted model. This establishes a proof of concept for knowledge-integrated discovery in sparse, highly structured materials domains.

Assembly modeling has been achieved in knowledge AI systems for the automated inference of new and rational metal–organic polyhedra (J. Am. Chem. Soc. 2022, 144, 26, 11713–11728). This work presents an algorithm and data structure that extends the process of assembly modeling to the automated generation of structural information about metal–organic polyhedra, enabling automation of computational approaches to analyse trends in cavity and pore sizing. Distinct from string‐based tools for purely organic cages, the workflow positions inorganic, organic, and hybrid chemical building units directly in 3‐D and outputs geometries suitable for higher‐level geometry optimization calculations in one step. The structural geometries obtained from this work are semantically integrated as part of The World Avatar, a dynamic knowledge ecosystem.

This work presents a generalisable process that transforms unstructured synthesis descriptions of metal–organic polyhedra into machine-readable, structured representations, integrating them into The World Avatar, a universal knowledge framework.

The open-source Python package ‘twa’ streamlines knowledge-graph-driven chemical data integration and workflow automation, unifying fragmented data and expanding the OntoMOPs knowledge graph with 799 new MOPs. It is available via ‘pip install twa’.

Metal–organic polyhedra (MOPs) are discrete, porous metal–organic assemblies known for their wide-ranging applications in separation, drug delivery, and catalysis. As part of The World Avatar (TWA) project—a universal and interoperable knowledge model—we have previously systematized known MOPs and expanded the explorable MOP space with novel targets. Although these data are available via a complex query language, a more user-friendly interface is desirable to enhance accessibility. To address a similar challenge in other chemistry domains, the natural language question-answering system “Marie” has been developed; however, its scalability is limited due to its reliance on supervised fine-tuning, which hinders its adaptability to new knowledge domains. In this article, we introduce an enhanced database of MOPs and a first-of-its-kind question-answering system tailored for MOP chemistry. By augmenting TWA’s MOP database with geometry data, we enable the visualization of not just empirically verified MOP structures but also machine-predicted ones. In addition, we renovated Marie’s semantic parser to adopt in-context few-shot learning, allowing seamless interaction with TWA’s extensive MOP repository. These advancements significantly improve the accessibility and versatility of TWA, marking an important step toward accelerating and automating the development of reticular materials with the aid of digital assistants.

Liquid formulations are ubiquitous yet have lengthy product development cycles owing to the complex physical interactions between ingredients making it difficult to tune formulations to customer-defined property targets. Interpolative ML models can accelerate liquid formulations design but are typically trained on limited sets of ingredients and without any structural information, which limits their out-of-training predictive capacity. To address this challenge, we selected eighteen formulation ingredients covering a diverse chemical space to prepare an open experimental dataset for training ML models for rinse-off formulations development. The resulting design space has an over 50-fold increase in dimensionality compared to our previous work. Here, we present a dataset of 812 formulations, including 294 stable samples, which cover the entire design space, with phase stability, turbidity, and high-fidelity rheology measurements generated on our semi-automated, ML-driven liquid formulations workflow. Our dataset has the unique attribute of sample-specific uncertainty measurements to train predictive surrogate models.