Artificial intelligence (AI) accelerates materials discovery, yet human expertise and education remain central to responsible, sustainable innovation.
New materials serve as the foundation for major technological progress, providing critical advancements in areas such as next-generation electronics, robotics, and medical devices. Traditionally, their development has relied on labour-intensive trial-and-error studies that are costly and slow. Moreover, the pace of new materials discovery is hindered by the vast design space. As a consequence, the average timeline for translating a new material from initial concept to commercial product is typically 10-20 years.¹
How AI is accelerating the pace of research and discovery
AI methods can now predict, discover, and optimise materials with far greater speed and efficiency. For example, Google’s Graph Networks for Materials Exploration (GNoME) deep learning tool predicted 2.2 million new crystals, identifying around 380,000 as stable materials.² Already, 736 of these have been synthesised by researchers, validating the AI’s predictive power. Moreover, AI-powered autonomous synthesis systems were able to create 41 novel compounds in just 17 days.³
Structural Constraint Integration in a GENerative model (SCIGEN) generated over 10 million candidate materials with specific lattice structures linked to quantum properties, from which 1 million passed stability screenings.⁴ Two novel compounds, TiPd0.22Bi0.88 and Ti0.5Pd1.5Sb, were synthesised and confirmed to exhibit paramagnetic and diamagnetic behaviour, validating the AI’s capability to bridge computational design with experimental reality. However, accelerating discovery is only the first step toward innovation. The translation from AI predictions to manufacturable materials still depends on processing, fabrication, and economic analysis that require expert judgment and cross-disciplinary co-ordination.
While AI speeds up discovery, these advances must be anchored in the atomistic understanding provided by physics-based material simulation. Traditional Density Functional Theory (DFT) and Molecular Dynamics (MD) simulations, while powerful for explaining and predicting atomic-level material properties, are computationally expensive and limited in scale. Machine-learned interatomic potentials (MLIPs) trained on massive datasets, such as OMol25, which contains over 100 million DFT evaluations spanning ~83 million unique molecular systems, can achieve near-DFT accuracy while dramatically reducing computational cost.⁵ Optimised frameworks can realise speedups in throughput for MD tasks, further narrowing the gap between high-accuracy simulation and practical usability. However, achieving DFT-level accuracy across complex systems remains an active research challenge. Together, AI-enhanced DFT and MD are redefining atomistic modelling – accelerating exploration from thousands to millions of candidate structures across energy, catalysis, battery, and biomaterials domains.
AI can close the loop between simulations and experiment through fully autonomous, data-driven workflows. Autonomous experimentation and self-driving labs (SDL) describe a system where AI, robotics, and automated processes work together in a closed-loop system to accelerate scientific research. For instance, a recent dynamic-flow SDL captured ten times more high-resolution reaction data at record speed, allowing to pinpoint promising inorganic materials in a single pass, greatly reducing the total number of experiments and cutting both time and material waste.⁶ Another advance shows self-supervised robotic systems mapping semiconductor properties: over 24 hours, the system autonomously drove a probe across 3,025 predicted points, enabling high-throughput, high-precision spatial characterisation.⁷ These examples show how SDLs can transform discovery timelines, making experimentation faster, smarter, and more resource-efficient.
Human experience and AI capability: A combined approach
While these successes are impressive, they underscore an equally critical reality: AI alone cannot replace the nuanced judgment and deep scientific intuition of human experts. While algorithms excel at generating large volumes of candidate materials and predicting performance metrics, only experienced researchers can rigorously evaluate the feasibility of synthesis, physical and chemical principles, scalability to industrial volumes, safety considerations, and long-term environmental sustainability. The most effective AI-augmented discovery teams are those where domain specialists apply their scientific intuition to filter AI-generated candidates, avoiding costly dead ends and steering research toward transformative innovations. This balance defines human-in-the-loop innovation, where AI accelerates and humans interpret, guide, and safeguard discoveries.⁸ This symbiotic relationship between human insight and AI capability will be pivotal to realising the full potential of AI-driven materials discovery and ensuring developments translate into impactful, viable technologies.
Education and training
If human judgment is central to responsible AI, then preparing the next generation with technical skills is essential. The economic demand for materials engineers with AI expertise is surging. Companies across energy, aerospace, electronics, and manufacturing are aggressively seeking engineers who can design experiments, interpret results, and accelerate innovation using AI tools and machine learning. Given that nearly every future materials engineer will engage with AI-enhanced, data-rich workflows, embedding AI literacy into materials science education is no longer optional; it is essential. Universities are experimenting with hybrid curricula that integrate AI modules into core science and engineering courses. Hands-on capstone projects, crash courses, and ‘data bootcamps’ are increasingly used to teach not only what AI can do, but what it should do.⁹ By shaping researchers into innovators who can both harness and question AI, education ensures technology serves humanity rather than the reverse.
Responsible AI for a thriving future
Overall, AI is no longer a distant promise but a driving force in materials science. It shortens discovery timelines, enables sustainable design, and integrates manufacturing through digital twins and adaptive materials. However, key challenges persist, such as issues with data and predictions quality, the interpretability of AI models, and the limited number of AI-trained researchers. High-quality, standardised materials datasets remain scarce, with many databases incomplete, inconsistent, or restricted, limiting robust model training and transferability. The complexity and context-dependence of materials, where performance often hinges on processing conditions and microstructure, further hinder generalisation. The lack of interpretability of many ML models reduces trust and makes it difficult to integrate predictions with physics-based frameworks, while the interdisciplinary expertise needed to develop and validate AI platforms limits adoption. Equally important are the unresolved limitations of AI in capturing processing-structure–property relationships and manufacturing feasibility. Most models optimise for thermodynamic stability or target properties without considering processing, production constraints, costs, or supply chain volatility, necessitating human-in-the-loop assessment and adaptive feedback loops to align AI predictions with real-world viability. These barriers underscore why AI’s success ultimately depends on human expertise. However, guided by human creativity and responsibility, AI can unlock materials and technologies that are not only faster and smarter but also transformative for society.
References
- Nekuda Malik JA. US National Academies report on the Frontiers of Materials Research. MRS Bulletin. 2019;44(5):329-334
- Merchant, A.; Batzner, S.; Schoenholz, S. S.; Aykol, M.; Cheon, G.; Cubuk, E. D. Scaling deep learning for materials discovery. Nature 2023, 624 (7990), 80-85
- Szymanski, N. J.; Rendy, B.; Fei, Y.; Kumar, R. E.; He, T.; Milsted, D.; McDermott, M. J.; Gallant, M.; Cubuk, E. D.; Merchant, A.; et al. An autonomous laboratory for the accelerated synthesis of novel materials. Nature 2023, 624 (7990), 86-91
- Okabe, R., Cheng, M., Chotrattanapituk, A. et al. Structural constraint integration in a generative model for the discovery of quantum materials. Nat. Mater. (2025)
- Levine, D. S. et al. “The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models.” arXiv preprint arXiv:2505.08762 (2025)
- Delgado-Licona, F., Alsaiari, A., Dickerson, H. et al. Flow-driven data intensification to accelerate autonomous inorganic materials discovery. Nat Chem Eng 2, 436–446 (2025)
- A. E. Siemenn et al. A self-supervised robotic system for autonomous contact-based spatial mapping of semiconductor properties. Sci. Adv 11,eadw7071(2025)
- Ramprasad, R., Batra, R., Pilania, G. et al. Machine learning in materials informatics: recent applications and prospects. npj Comput Mater 3, 54 (2017)
- T. J. Oweida, A. Ul-Mahmood, M. D. Manning, S. Rigin, Y. G. Yingling, “Merging Materials and Data Science: Opportunities, Challenges, and Education in Materials Informatics”, MRS Advances 5 (2020) 1-18
Please note, this article will also appear in the 24th edition of our quarterly publication.
