Researchers Develop Integrated Workflows for Reliable Materials Discovery

A thorough combination of theoretical modelling, computational analysis, experimentation, and machine learning is currently aiding materials discovery. Phalgun Lolur and colleagues from Capgemini Quantum Lab, Research Triangle Park, King’s College London Strand, University of Nottingham, PsiQuantum, Daresbury Laboratory, Toyota Research Institute, Fraunhofer Institute for Mechanics of Materials IWM, Robert Bosch GmbH, Robert-Bosch-Campus 1, Airbus Central Research and Technology, Lawrence Berkeley National Laboratory, GE Vernova Advanced Research, outline the current state-of-the-art, limitations, and future opportunities in this field. Their work highlights the key role of reproducible workflows and shared data standards in delivering strong predictions and accelerating the development of deployable materials, bridging the gap between initial design and practical application. This integrated approach enhances decision-making and streamlines the materials’ discovery process by combining the broad coverage of classical simulations with insights from experimental measurements and the potential of quantum computing for understanding correlated electronic behaviour. The increasing complexity of materials science necessitates a holistic approach, moving beyond single-method investigations to leverage the strengths of each discipline and mitigate individual weaknesses. Historically, materials discovery relied heavily on serendipity and empirical trial-and-error, a process that is both time-consuming and resource-intensive. Modern computational methods offer the potential to significantly accelerate this process, but their effectiveness is contingent upon accurate modelling of material behaviour and reliable prediction of properties.

Quantifying computational variance to identify consistently high-performing materials

Discrepancies between computational approximations now reach tenths of an electron volt, a threshold previously limiting the reliable ranking of candidate materials for advanced applications. Such disagreement, sufficient to alter predicted material performance, previously necessitated excessively conservative designs or hindered accurate comparison between theoretical predictions and experimental results. These variations stem from the inherent approximations within computational methods, such as density functional theory (DFT), where the exchange-correlation functional introduces uncertainty. Different functionals, even within the same DFT framework, can yield significantly different results for material properties. Systematic workflows reveal these variations, enabling a statistical understanding of uncertainty and prioritising “strong optima”, materials performing consistently under realistic conditions, rather than solely focusing on theoretical ideals. This statistical approach is crucial for moving beyond simply identifying promising candidates to confidently predicting their behaviour in real-world scenarios. The concept of ‘strong optima’ acknowledges that a material performing consistently well across a range of conditions is often more valuable than one exhibiting exceptional, but potentially fragile, performance under idealised circumstances.

National laboratories are establishing integrated ecosystems where simulation, analysis, and data management function cohesively, rather than as isolated processes. This coordinated approach extends beyond simply improving numerical precision, prioritising advancements that demonstrably enhance decision-making in materials selection. Facility-scale computing and software modernisation are broadening access to high-fidelity modelling, allowing more groups to validate results and establish shared benchmarks reflecting genuine materials challenges. Algorithms are now evaluated not solely on their accuracy in reproducing known values, but on their effectiveness in guiding exploration and mitigating risk, with some codes now accessible to a wider audience thanks to adaptations for heterogeneous hardware and exascale architectures. The development of standardised data formats and ontologies is also critical, enabling seamless data exchange and facilitating the integration of data from diverse sources. This interoperability is essential for building comprehensive materials databases and accelerating the development of machine learning models.

Computational Variation and Statistical Assessment of Materials Properties

Uncertainty quantification, a technique for estimating potential error in calculations, became central to this work by acknowledging the limits of predictive power. Researchers developed workflows that systematically varied the computational methods used to model materials, running multiple simulations with differing levels of approximation. Comparing results from these diverse calculations built a statistical picture of the uncertainty surrounding each material’s predicted properties, much like acknowledging a margin of error in opinion polls. This process involves exploring the ‘computational landscape’ of a material, identifying regions where predictions are robust and less sensitive to methodological choices. Techniques such as Bayesian optimisation and Gaussian process regression are increasingly employed to efficiently explore this landscape and quantify uncertainty. The choice of computational parameters, such as the k-point mesh density in DFT calculations, also contributes to uncertainty and must be systematically varied.

Workflows focused on the complexity of real systems, integrating theory, computation, experiment, and machine learning. Classical simulations were used for broad material screening, while experimental measurements assessed degradation and performance under realistic conditions. Machine learning accelerated exploration using curated datasets, and this approach prioritised identifying “strong optima”, materials maintaining performance despite real-world variations, over pursuing theoretical “global optima”. Experimental validation is crucial for calibrating computational models and ensuring their predictive accuracy. Techniques such as in-situ characterisation, where materials are studied under operating conditions, provide valuable insights into their behaviour and can help to refine computational models. The integration of machine learning allows for the efficient analysis of large datasets and the identification of patterns that might be missed by traditional methods.

Embracing imperfection delivers strong materials for real-world applications

Reliable materials depend on workflows that mirror the messy reality of how things break down and behave over time. Computational power and machine learning accelerate discovery, yet a persistent tension exists between seeking theoretical perfection and achieving consistent performance in the real world. This gap arises because current methods often prioritise identifying the absolute “best” material, rather than one that reliably functions within the inevitable imperfections of manufacturing and use. Factors such as defects, grain boundaries, and surface effects can significantly influence material behaviour, and these are often difficult to accurately model in simulations. Furthermore, manufacturing processes introduce inherent variations that can affect material properties.

Acknowledging the inherent difficulty in perfectly predicting real-world material behaviour does not justify abandoning these advanced workflows. The team’s approach prioritises building systems that account for imperfections from the outset, rather than striving for an unattainable ideal. This focus on durability, even if it means sacrificing the absolute “best” theoretical material, delivers more consistently functional products and accelerates the path to deployment. This requires a shift in mindset, from seeking the ‘perfect’ material to designing materials that are robust and resilient to imperfections. Incorporating realistic defect concentrations and surface conditions into simulations is crucial for achieving this goal.

Systems are being built that prioritise function over theoretical perfection, embracing imperfections from the start to enhance durability. This delivers consistently functional products and will reshape materials discovery for a new era of strong design. Prioritising reliably performing materials, even if not theoretically perfect, accelerates practical application. This shift demands workflows integrating modelling, experimentation, and data analysis, acknowledging that discrepancies in calculations can alter material rankings. By systematically assessing uncertainty, scientists can move beyond simply identifying potential materials to confidently predicting their behaviour under real-world conditions, opening questions regarding the development of standardised benchmarks reflecting operational complexities and enabling more accurate validation of computational methods and predictive power. The development of such benchmarks will be essential for fostering trust in computational predictions and accelerating the adoption of new materials in critical applications, such as energy storage and aerospace engineering.

The research demonstrated that integrating computational modelling, experimentation, and data analysis improves materials discovery workflows. Understanding and accounting for imperfections, such as manufacturing variations and realistic defect concentrations, is crucial, as striving for theoretical perfection does not guarantee functional materials. This approach prioritises designing robust and resilient materials that perform consistently under real-world conditions, rather than identifying solely the ‘best’ theoretical options. The authors suggest developing standardised benchmarks to assess uncertainty and validate computational predictions, ultimately fostering confidence in material behaviour.