SWITCHED ON
The daily technology series nobody asked for but everyone needed
The Materials Machine: AI and the Compressed Timeline of Discovery
Humanity has discovered roughly 20,000 stable inorganic materials over the entire history of chemistry. In 2023, a single DeepMind model predicted 2.2 million more in one project. The gap between predicting a material and actually making and using one is where this story gets interesting.
Materials science has historically progressed through a combination of accident, intuition, and exhaustive trial and error — the transistor, the lithium-ion battery, and most of the materials underlying modern technology were discovered through decades of patient, largely manual experimentation. AI is compressing the search space of that process dramatically. It has not yet compressed the much harder problem of actually synthesising and validating what the search finds.
Yesterday we mapped the geopolitics of critical minerals — China's processing chokepoint, the DRC's cobalt dilemma, the Western response's realistic decade-plus timeline, and recycling as the underrated structurally self-sufficient long-term answer. Today we are looking at a technology that could, if it delivers on its early promise, meaningfully change several of the resource and technology constraints this series has documented across three seasons: AI-driven materials discovery. The same fundamental capability that produced AlphaFold's protein structure breakthrough, which we covered in Season One, has been redirected toward an even larger and in some ways more commercially consequential target — discovering entirely new materials, from battery chemistries to superconductors to catalysts, at a pace that conventional experimental materials science cannot match.
01 — Why Materials Discovery Was So Slow
Finding a new material with useful properties has historically been an extraordinarily inefficient process, for reasons that are instructive to understand before appreciating what AI changes. The space of theoretically possible inorganic compounds — different combinations of elements in different ratios and crystal structures — is vast, and the properties that emerge from any specific combination (electrical conductivity, mechanical strength, thermal stability, chemical reactivity) depend on complex quantum mechanical interactions that are extremely difficult to predict from first principles without immense computational resources.
The conventional approach has combined scientific intuition (chemists' accumulated understanding of which combinations are likely to be interesting, built from decades of training and experience), incremental variation on known materials (adjusting the ratios or substituting elements in an existing compound with known useful properties), and a great deal of laborious trial-and-error laboratory synthesis and testing. This process has produced genuinely transformative discoveries — silicon semiconductors, lithium-ion battery chemistries, high-temperature superconductors — but at a pace measured in years to decades per major discovery, constrained fundamentally by how much experimental testing a human research team can physically perform.
02 — GNoME and the Scale of the Prediction Problem
Google DeepMind's GNoME (Graph Networks for Materials Exploration) project, published in Nature in November 2023, represents the most dramatic demonstration to date of AI's capability to compress the materials prediction problem. Using graph neural networks trained on known crystal structures and their properties, combined with active learning techniques that iteratively improve the model's predictions, GNoME predicted approximately 2.2 million new stable crystal structures — of which roughly 380,000 were identified as the most promising candidates for further investigation, representing a roughly order-of-magnitude expansion of the total number of known stable inorganic materials in a single research project, compared to the approximately 48,000 stable materials that had been discovered and catalogued through all of human materials science history to that point.
The scale of this prediction is genuinely without precedent, and it is important to understand precisely what it does and does not represent. GNoME predicts which crystal structures are thermodynamically stable and therefore could, in principle, exist as real materials — a computational filtering step that dramatically narrows the space experimental researchers need to explore, from the astronomically large space of theoretically conceivable combinations down to a still-large but far more tractable set of predicted-stable candidates. It does not synthesise these materials, verify their real-world properties beyond the computational prediction, or determine whether they can be manufactured at practical cost and scale. The prediction is the first, and historically most time-consuming, step in a much longer discovery pipeline.
Predicting that a material is theoretically stable is analogous to identifying that a chess position is theoretically reachable. It tells you the move sequence is legal. It tells you nothing about whether the resulting material is useful, manufacturable, or better than what already exists for any specific application — questions that still require the harder, slower work of actual laboratory validation.
03 — Self-Driving Labs and Closing the Validation Gap
The most significant complementary development to AI prediction models is the emergence of automated, robotic "self-driving laboratories" — physical laboratory systems where robotic equipment, guided by AI models, can synthesise candidate materials, run characterisation tests, and feed the results back into the prediction model to guide the next round of experiments, substantially compressing the traditionally manual and slow experimental validation stage that follows computational prediction.
A Berkeley Lab and University of California collaboration demonstrated an autonomous materials synthesis system, called the A-Lab, that combined AI-guided experiment design with robotic synthesis and characterisation equipment to synthesise 41 novel materials predicted by computational models in seventeen days of largely unsupervised operation — a pace that would have taken a human research team months to years using conventional manual synthesis and characterisation methods. Similar self-driving lab initiatives are underway at multiple research institutions and, increasingly, within industrial research and development organisations including major battery and semiconductor manufacturers, reflecting the commercial value of compressing the historically slow discovery-to-validation pipeline.
The honest limitation of self-driving labs as of 2026 is that they remain most effective for materials within specific, relatively well-understood chemical families where the synthesis routes and characterisation methods are already established and can be automated — genuinely novel synthesis challenges, requiring new chemical processes rather than variations on known ones, still require the kind of expert human chemical intuition that automated systems have not yet replicated. This mirrors a pattern this series has documented repeatedly in different domains: automation excels at well-specified, well-bounded problems and struggles more with genuinely novel ones requiring judgment under uncertainty.
04 — The Specific Materials That Matter Most
Several application domains illustrate where AI-driven materials discovery is generating the most commercially and strategically significant near-term interest. Solid-state battery electrolytes — materials that could replace the liquid electrolytes in current lithium-ion batteries with solid alternatives, potentially enabling higher energy density, faster charging, and improved safety by eliminating the flammability risk associated with liquid electrolytes — represent one of the most actively pursued applications, with several AI-identified candidate materials currently moving through industrial validation at battery manufacturers including those we discussed in the electric vehicles episode.
Carbon capture materials — porous frameworks called metal-organic frameworks that can selectively capture carbon dioxide from industrial exhaust or ambient air, relevant to the direct air capture technologies we examined critically in Season One's climate technology episode — have seen a substantial expansion of candidate materials identified through AI screening, with the practical challenge remaining the same one that constrained conventional direct air capture: even an optimal capture material must still be deployed and operated at the enormous physical scale required to meaningfully affect atmospheric carbon concentrations, a scale-up challenge that materials discovery alone does not solve. Superconductor research, seeking materials that conduct electricity without resistance at higher and more commercially practical temperatures than current superconductors (which typically require expensive cryogenic cooling), has also benefited from AI-guided screening, though no AI-predicted material has yet demonstrated the combination of high critical temperature and practical manufacturability that would represent the kind of breakthrough the field has sought for decades.
05 — Why the Compressed Timeline Still Takes Time
The honest assessment of AI materials discovery, consistent with the pattern this series applies to every technology it covers, is that the prediction and screening stage of materials discovery has genuinely and dramatically compressed — a process that took years of researcher time to narrow down promising candidates can now be substantially accelerated through computational screening. What has not compressed to nearly the same degree is the subsequent validation, scale-up, manufacturing process development, and cost optimisation required to take a promising predicted material from a laboratory curiosity to a commercially deployed product — stages that remain constrained by physical, chemical, and economic realities that AI prediction does not directly address.
This is the same pattern that appeared in our AlphaFold and drug discovery discussion in the AI-in-medicine episode: the AI compresses the discovery and design stage substantially, while the clinical trial, regulatory, and manufacturing stages that follow remain paced by processes AI has not yet meaningfully accelerated. The realistic expectation for AI-driven materials discovery's practical impact is not an overnight transformation of available materials, but a steadily accelerating pipeline of new materials reaching commercial validation over the coming decade, with the compressed discovery stage becoming visible in product terms gradually, as each individually promising candidate works its way through the still-slow subsequent stages of the materials development pipeline.
Tomorrow we are shifting to a topic with more immediate personal stakes — the technology and ethics of workplace surveillance, specifically the AI monitoring tools increasingly deployed to track remote and hybrid worker productivity, and what the evidence actually shows about their effectiveness and their costs. See you then.
Switched On is a daily technology series covering the ideas, systems, and arguments shaping the digital world. Opinionated. Witty. Occasionally wrong. Always worth the argument.
