Self-driving laboratories (SDLs) represent a transformative paradigm in chemical and materials discovery, integrating automated hardware with artificial intelligence to drastically accelerate research cycles. Realizing their full potential, however, requires moving away from fragile, laboratory-specific scripts toward a unified and robust digital infrastructure.

This presentation first outlines the fundamental architecture of automated laboratories, examining core design principles for constructing scalable, multi-disciplinary infrastructure. Building upon these principles, a minimal, language-agnostic orchestration foundation is introduced. This framework structures laboratory behavior into a four-level hierarchy: primitives, unit operations, state-preserving unit flows, and workflows. By establishing a rigorous typing discipline and state-preservation contracts, the architecture enables safe parallel execution and robust validation. Crucially, this layered abstraction democratizes workflow design, allowing complex experimental tasks to be efficiently distributed across different scientists, software engineers, and domain experts under a unified "Lab as Code" paradigm.

Finally, the discussion explores how databases and structured knowledge extracted from scientific literature can be seamlessly integrated into these automated ecosystems. By bridging the gap between historical literature data and autonomous execution, this approach provides a scalable foundation for continuous, data-driven scientific discovery.

Yttria-stabilised zirconia and gadolinium-doped ceria are key oxide ion conductors used in solid oxide fuel cells, and they are often combined in multilayer electrolyte architectures. During high-temperature fabrication and operation, cation interdiffusion can generate a mixed YSZ/GDC interdiffusion layer that is commonly associated with additional interfacial resistance. Classical molecular dynamics can provide useful atomistic insight into local cation chemistry and oxygen vacancy transport. However, empirical force fields developed separately, even when compatible and based on similar parametrisation strategies, may not provide a consistent common benchmark for comparing different oxide electrolytes, as illustrated by cases where they invert the experimentally established transport ordering between YSZ and GDC. This discrepancy motivates reference-based approaches that can bridge first-principles accuracy and molecular dynamics time and length scales.

Here, we develop a workflow to adapt foundation machine-learned interatomic potentials to YSZ, GDC, and mixed Ce, Zr, Y, Gd oxide compositions. Charge-neutral fluorite models were constructed to represent bulk-like endmembers and YSZ/GDC mixed compositions with controlled cation and oxygen vacancy environments. A consistent first-principles reference dataset was generated from strained structures, randomly displaced configurations, finite-temperature snapshots, and oxygen vacancy hop environments. Foundation potentials were then fine-tuned using nested training subsets to quantify how much first-principles data is required to achieve reliable accuracy.

The resulting models are evaluated not only by energy and force errors, but also by oxygen mean-squared displacements, tracer diffusivities, Arrhenius behaviour, and vacancy hop energetics. The results indicate that oxygen transport depends strongly on the sampled cation and oxygen vacancy environments, and that improving the fit to first-principles data does not by itself guarantee recovery of experimental transport trends. Instead, the persistent discrepancy points to an upstream limitation in the reference description, defect chemistry, or sampled vacancy environments, rather than simply to insufficient model fitting accuracy. This work establishes a physics-grounded benchmark for foundation machine-learned interatomic potentials in oxide ion conductors and provides a route toward predictive modelling of transport limitations in realistic YSZ/GDC interdiffusion layers.

Predicting the three-dimensional conformation of molecules from their bonding graph is a fundamental problem in computational chemistry, with applications in drug discovery, materials science, and molecular simulation. We present a bond-centric graph neural network framework for predicting local molecular geometry directly from graph connectivity. Instead of generating full Cartesian coordinates in a single step, the method decomposes the problem into local predictions around individual bonds, using a graph attention network to encode chemical environments and a rotationally invariant spherical representation to model bond distances and angular structure. To support this formulation, we construct an independently generated benchmark dataset of 9,691 DFT-optimized molecular structures using a large-scale Quantum ESPRESSO workflow initialized from PubChemQC molecular candidates. The resulting dataset provides rotationally consistent local geometric supervision for bond-centric molecular reconstruction while remaining independent from the original molecular optimization pipeline. On the benchmark set, the proposed framework achieves a heavy-atom bond-distance mean absolute error of 0.103 Å and a root mean squared error of 0.152 Å, corresponding to an average relative bond-length error of approximately 7.37%. Local angular reconstruction achieves a polar-angle θ mean absolute error of 4.56° and root mean squared error of 7.49°, while azimuthal-angle Δφ prediction remains substantially more challenging, with a mean absolute error of 18.19° and root mean squared error of 31.34°. Error analysis shows that bond lengths and local bond-angle magnitudes are reconstructed reliably across diverse chemical environments, whereas torsional and rotational degrees of freedom contribute the dominant source of geometric uncertainty, particularly in flexible single-bond conformations. Qualitative reconstruction results demonstrate strong preservation of molecular topology and local bonding environments, with successful extrapolation to polymeric systems significantly larger than those observed during training. From a computational perspective, the method exhibits predictable scaling with system size, following a power-law relationship of approximately T proportional to B^1.98 with respect to bond count. In direct comparison, RDKit ETKDG-only and ETKDG + MMFF workflows exhibit steeper empirical scaling trends of approximately T proportional to B^3.04 and T proportional to B^2.99, respectively. These results indicate that the proposed bond-centric reconstruction strategy becomes increasingly advantageous for large polymer-scale molecular graphs. Overall, these results demonstrate that bond-centric local geometry prediction provides a scalable and physically consistent approach for rapid three-dimensional molecular structure generation, offering a practical alternative for high-throughput molecular modeling.

AI-driven autonomous laboratories have accelerated materials discovery by adapting conventional automated equipment like liquid handlers, robotic arms, spin coaters, to replace human-executed workflows. While these systems have successfully automated traditional experimentation, they remain constrained by a fundamental design limitation: chemical processes built around a human throughput model, where experiments run sequentially rather than in parallel.

We argue that this model of laboratory automation represents a ceiling, not just a bottleneck and we need to reimagine chemistry and chemical labs to build AI-native labs, which does not limit AI, but make use of its full potential to accelerate materials innovation. For. e.g. even though a batch Bayesian optimization proposes many experiments at once, they are still physically executed one by one due to laboratory constraints, with results fed back only after all the experiments have been executed. Unlike human chemists, AI can plan, execute and analyze parallel experimental streams, yet current automated laboratories are not utilizing this capacity of AI to its full potential.

We challenged ourselves to address this challenge and build an AI-native physical laboratory architected from the ground up for AI execution, not human adaptation. Rather than automating human processes, HELIOS reimagines chemical synthesis as a massively parallel, AI-directed operation capable of running one million material experiments per minute. As a proof of concept, we apply HELIOS to the exploration of the vast perovskite chemical space for energy applications, demonstrating that AI-native lab design can unlock a scale of experimentation previously inaccessible to materials science.

The discovery of sustainable materials requires navigating vast chemical and processing spaces while balancing performance, stability, scalability, and resource constraints. Conventional trial-and-error experimentation is often too slow and data-limited to meet the urgency of challenges in clean energy, sustainable manufacturing, and next-generation electronics. In this talk, I will describe how self-driving laboratories that combine robotics, automated synthesis and characterization, and machine-learning-guided decision-making can transform materials discovery into a data-rich, closed-loop process. I will highlight recent work from our group on autonomous platforms for semiconductor nanomaterials, halide perovskite quantum dots, and catalytic materials, where AI-guided experimentation accelerates optimization, reveals hidden synthesis–structure–property relationships, and reduces waste generation. Emphasis will be placed on how different modalities of self-driving laboratories, including flow-based, batch-based, and multi-robot platforms, can support sustainable materials development by increasing experimental throughput, improving reproducibility, and enabling rapid learning from every experiment. These advances point toward a future in which autonomous laboratories serve as broadly accessible engines for discovering and manufacturing high-performance materials with lower environmental and economic cost.

The discovery of functional nanomaterials for sustainable technologies is challenged by vast multidimensional chemical spaces and by the difficulty of closing the loop between synthesis, characterization, data interpretation, and decision-making. This challenge is particularly acute for nanoparticles used in catalysis and energy-related applications, because their properties and performance depend sensitively on chemical composition, atomic structure, and nanoparticle size and shape.

What a self-driving laboratory can learn ultimately depends on the information it is allowed to learn from. In this context, information fidelity reflects both data quality and relevance to the target property or structure.

When total X-ray scattering is combined with atomic pair distribution function (TXS-PDF), the reciprocal-space total scattering data are Fourier transformed into real-space atomic pair correlations, enabling local atomic structure to be probed even when nanoparticles are too small or disordered for conventional diffraction analysis [1]. 
High-fidelity TXS-PDF data are typically acquired at synchrotron X-ray sources, where high-brilliance X-ray beams enable rapid acquisition of structure-sensitive data. ScatterLab has so far been established as a synchrotron-based self-driving laboratory [2]. Unfortunately, synchrotron access is limited. This motivates a transition from synchrotron-based operation toward in-house workflows using laboratory X-ray sources.

While TXS-PDF is powerful for atomic structure characterization, small-angle X-ray scattering (SAXS) provides a more direct and robust probe for particle size and shape [3]. Together with rapid optical probes such as UV-Vis spectroscopy, TXS-PDF and SAXS provide complementary information on optical response, atomic structure, and nanoscale morphology.

One of the greatest assets of ScatterLab is our robotic modular experimentation platform (MODEX). MODEX automates solution-based chemical synthesis through modular unit operations such as dispensing, mixing, illumination, and sample transfer. This modularity makes the platform configurable across synthesis protocols and transferable between characterization instruments.

Bayesian optimization turns ScatterLab from an automated platform into an autonomous laboratory by selecting the next synthesis parameters based on previous experimental outcomes. In our current implementation, the optimization target is the agreement between experimental and simulated scattering patterns, quantified using the mean squared error.

ScatterLab uses robotics and Bayesian optimization to close the loop around structure-sensitive feedback. The next challenge is moving from scarce synchrotron measurements to in-house multimodal characterization, where the autonomous laboratory learns from fit-for-purpose information relevant to functional nanomaterials for catalysis and energy applications.

The development of high efficiency perovskite solar cells relies heavily on identifying the precise combination of processing conditions and material compositions. Poor reproducibility and sensitivity to processing parameters have made it difficult to systematically optimize perovskite thin films across labs, and the high dimensionality of the processing space leaves traditional trial and error methods poorly equipped to navigate the search space. Self-driving laboratories offer a route to make this optimization more efficient and effective by coupling robotic synthesis with machine learning-driven decision making. Here we present a fully automated, closed loop spin coating platform driven by high dimensional Bayesian optimization to autonomously synthesize and optimize perovskite thin films.

We employ the SpinBot platform with custom modifications for closed loop integration. The system autonomously performs solution mixing, spin coating, thermal annealing, and characterization. It is capable of fabricating nearly the entire device stack on platform, with only electrode evaporation performed externally. Because completing and testing full devices is comparatively slow and would bottleneck the optimization loop, we instead characterize each film in-line with a suite of optical techniques; PL and absorption mapping, time-resolved PL, and intensity-modulated PL spectroscopy that probe the carrier dynamics and ionic transport governing device efficiency and stability. A custom software orchestration framework bridges the robotic hardware with the optimization loop, enabling continuous experimental iteration. As an initial demonstration, we used the platform to run an optimization of 20 processing parameters over 300 samples on a single perovskite composition, confirming that the integrated synthesis-characterization-optimization loop operates reliably and converges to improved film quality.

Building on this validated loop, we now target a more device relevant objective. From the combined characterization data, we construct a proxy metric for device efficiency and stability, and optimize processing conditions against it in a closed loop using Bayesian optimization. The optimization spans 10-20 continuous and categorical processing parameters, including antisolvent type, timing, and flow rate, alongside spin-coating and annealing parameters, and extends to precursor ink composition through in-situ solution mixing. Constructing this proxy from fast optical measurements provides the throughput needed to navigate such a large parameter space.

Self-driving laboratories (SDLs) promise to fundamentally transform how we discover and scale advanced materials, yet realising that promise demands more than automating existing workflows. This talk presents our group’s expanding portfolio of SDL platforms and illustrates how the paradigm generalises across vastly different materials challenges. We begin with our flagship photovoltaic work, in which an SDL-enabled roll-to-roll line fabricated perovskite solar cell modules at 11% efficiency under ambient conditions, processing over 11,800 devices in 24 hours at a projected cost of approximately 0.7 USD/W. Building on the lessons learned, we describe HYDRA, our recently funded NRF AI for Science programme that deploys a closed-loop SDL for e-waste hydrometallurgy, coupling an Opentrons Flex liquid-handling robot with Agilent ICP-OES analytics to autonomously optimise metal-recovery processes from electronic waste. We then turn to an emerging frontier: AI-guided scaffold design for cultivated meat, where data-driven optimisation of biomaterial composition and architecture is being pursued under a dedicated SFA grant. Across these domains, a common thread emerges: the evolution of SDLs from narrow optimisation engines toward autonomous systems capable of hypothesis generation, cross-domain transfer, and integration of manufacturing constraints from the outset. We discuss how this vision is supported by our broader work in AI-driven materials discovery, including the development of autonomous agents and open-source tools that aim to democratise SDL access. The talk concludes with a roadmap for next-generation SDLs and reflections on the translational pathway from laboratory discovery to commercial impact, drawing on our ongoing technology-transfer and commercialisation efforts.

Hydrogen fuel plays a critical role for the transition towards a sustainable future with utilisation in energy, chemical and other industrial sectors. Proton exchange membrane (PEM) electrolysers have been highlighted by the UK government as the most suitable technology for green hydrogen production as upon scale-up, there is a greater potential for cost reductions with increased efficiency.¹ However, PEM electrolysers employ expensive and scarce iridium oxide electrocatalysts for the oxygen evolution reaction (OER) which generates a bottleneck towards its commercialisation.² Futhermore, finding a replacement catalyst material is restricted due to the harsh operational conditions and acidic environment.³

Here we utilise automated, high-throughput (HT) synthesis, testing and characterisation methods to allow for facile screening and optimisation of potential electrocatalysts for OER in PEM electrolysis. Adam’s Fusion synthesis is performed in a Chemspeed FLEX ISYNTH and FLEX CATSCREEN to synthesise and anneal 24 catalysts in one day. In this work, we concluded a 3-fold decrease of reagent sodium nitrate achieves the same electrochemical performance compared to literature, and annealing at 450 ^oC for 1 h provides good OER performance. With XRD and TEM, we showed that these heat treatment conditions generate a catalyst with a long-range amorphous but short-range rutile crystal structure – good indications of both an active and stable OER electrocatalyst. Exploration of acid washes too has doubled the electrochemical surface area of catalyst, matching the activity of commercial iridium oxide catalysts. In turn, automated and HT electrochemical testing (DOBOT, Opentrons, multi working electrode (MWE)) are implemented to allow for seamless ink production and electrochemical assessment of materials produced. With the multi working electrode, six electrocatalysts can be tested at once compared to a tradition rotating disc electrode for facile activity screening. Moreover, initial DFT screening has highlighted potential rutile iridium mixed oxide catalysts, in which the results have been utilised towards a Bayesian optimisation campaign. Overall, this workflow allows for facile screening and optimisation of low-iridium mixed oxide catalysts, underlining active and stable OER catalysts of various compositions.

Though the work is preliminary – and focuses on the benchmarking of pure iridium oxide synthesis – it highlights advantages of automated, high-throughput workflows including speed and reliability due to lack of human involvement, screening larger syntheses spaces at once, and reduced reagent consumption, aligning with principles of green chemistry.

Materials Acceleration Platforms (MAPs) are emerging as a powerful paradigm for closing the loop between materials synthesis, characterization, data analysis, and decision-making, enabling faster and more autonomous materials discovery. In this talk, we present an electrochemistry-focused MAP developed for the accelerated design of functional interfaces relevant to energy conversion, electrocatalysis, and corrosion protection. The platform integrates automated liquid handling, parallelized electrochemical experimentation, standardized workflows, and machine learning-driven experiment planning within a unified digital infrastructure.

Central to the approach is the generation of high-quality, FAIR-compatible datasets with structured metadata that support active learning, inverse design, and reproducible experimentation. Electrochemical descriptors are employed as application-specific proxy metrics and correlated with long-term performance indicators, enabling rapid identification of promising materials while reducing experimental burden. To efficiently navigate complex design spaces, we combine Bayesian optimization, Pareto-based multi-objective analysis, and alternative machine-learning strategies, highlighting how algorithm performance depends strongly on the underlying materials system and optimization objective.

Through case studies on electrodeposited metallic and alloy thin films, we demonstrate accelerated optimization of corrosion resistance and catalytic activity, while uncovering processing-structure-property relationships that guide materials design. The presentation will summarize the design and construction phases of our MAPs, their constituent modules, and workflows. It will also include deep dives into examples on Pareto optimization and evaluation of proxy electrochemical experiments for MAP-deployment. Beyond individual applications, the presentation will also discuss practical challenges that remain critical for realizing fully autonomous discovery workflows, including instrument interoperability and integration of legacy characterization techniques.