What can AI actually do at scale today?
AI can now screen millions of candidate materials computationally, a task that was previously impractical. In 2023, researchers trained a graph neural network on 48,000 known stable crystals and used it to predict 2.2 million new stable structures below the current convex hull—a measure of thermodynamic stability. That is an order-of-magnitude increase in the number of stable materials known to humanity [2]. Of those predicted structures, 736 had already been independently synthesized by other labs, providing real-world validation that the AI's predictions are not just theoretical [2].
A 2024 study took a different approach, combining machine learning models with traditional physics-based calculations on cloud high-performance computing (HPC) to screen over 32 million candidate materials for solid-state battery electrolytes. The pipeline identified roughly half a million potentially stable materials, then narrowed those down to 18 promising new compositions. One of those, the NaxLi3-xYCl6 series, was synthesized and experimentally characterized, confirming its potential as a solid electrolyte [4]. This shows that AI can not only predict but also guide experimental synthesis in a targeted way.
Where does AI still fall short?
Despite these successes, a 2024 critical analysis of the same Google-led work found that many of the AI-predicted compounds lack novelty, credibility, or practical utility. The authors argue that the methods hold promise but that there is a 'great need to incorporate domain expertise in materials synthesis and crystallography' [1]. In other words, AI can generate plausible candidates, but turning those candidates into real, useful materials still requires human judgment and hands-on lab work.
The bottleneck has shifted from prediction to validation. While AI can propose millions of candidates, only a tiny fraction get experimentally tested. In the battery electrolyte study, only one new material series was fully synthesized and characterized out of 18 top candidates [4]. Similarly, a 2025 study used an AI search engine to discover 50 new altermagnetic materials—a rare magnetic phase—but all were confirmed only through first-principles electronic structure calculations, not actual synthesis [5]. Experimental validation remains the rate-limiting step.
Another challenge is that AI models are only as good as the data they are trained on. Many existing materials databases are sparse, biased toward well-studied chemistries, or lack information on synthesis conditions. A 2022 review on machine learning for thermal energy materials noted that the field is 'still in the early stage' and that data quality and feature engineering are critical hurdles [7]. Without high-quality, diverse training data, AI predictions can miss entire classes of materials or overfit to known patterns.
Is industry actually using this yet?
Yes, but mostly in targeted applications rather than as a general-purpose tool. The U.S. Department of Energy has identified AI-driven materials design as pivotal for applications ranging from solar cells and batteries to nuclear fuels and carbon capture [6]. Self-driving labs—which combine AI, robotics, and automated experimentation—are being developed to accelerate the discovery cycle by 10 to 100 times, though they are still in the prototype stage [8].
In the battery industry, AI is already being used to screen redox flow battery materials by integrating high-throughput computational screening with machine learning, reinforcement learning, and Bayesian optimization [10]. Similarly, in optoelectronics, machine learning is speeding up ab initio calculations and guiding experimental synthesis of perovskites and organic semiconductors [9]. These are not yet fully automated pipelines, but they represent real industrial engagement with AI-guided discovery.
The 'Safe and Sustainable by Design' framework is also incorporating AI to predict toxicity and lifecycle impacts before materials are produced, which could help industry avoid costly late-stage failures [3]. This suggests that AI's role is expanding beyond discovery into the broader product development lifecycle.
About These Sources
This answer is built on 10 peer-reviewed studies — published from 2022 to 2026, 6 from 2024 or later, 8 in Q1 journals, collectively cited 1,350 times — selected as the most relevant from 10 studies that passed quality screening, drawn from 53 papers retrieved from a database of over 500 million.
Sources used in this answer
Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery
A critical perspective on the Google-led deep learning study found scant evidence that AI-predicted compounds meet the trifecta of novelty, credibility, and utility, emphasizing the need for domain expertise in synthesis and crystallography.
Scaling deep learning for materials discovery
Graph networks trained on 48,000 stable crystals predicted 2.2 million new stable materials—an order-of-magnitude expansion—and 736 of those predictions have been independently experimentally realized.
Safe and sustainable by design with ML/AI: A transformative approach to advancing nanotechnology
Machine learning and AI are enhancing the 'Safe and Sustainable by Design' framework for nanotechnology by enabling predictive toxicology, materials informatics, and lifecycle analysis.
Accelerating Computational Materials Discovery with Machine Learning and Cloud High-Performance Computing: from Large-Scale Screening to Experimental Validation
Combining ML and cloud HPC screened 32 million candidates for solid-state battery electrolytes, identifying 18 promising compositions; one series (NaxLi3-xYCl6) was synthesized and experimentally validated.
AI-accelerated discovery of altermagnetic materials
An AI search engine using a pre-trained graph neural network discovered 50 new altermagnetic materials (metals, semiconductors, insulators), including four i-wave altermagnets, confirmed by first-principles calculations.
AI for Materials Design and Discovery Using Atomistic Scale Information [Industrial and Governmental Activities]
The U.S. Department of Energy identifies AI-driven materials design as pivotal for applications in renewable energy, energy storage, carbon capture, and nuclear energy.
Machine Learning for Harnessing Thermal Energy: From Materials Discovery to System Optimization
Machine learning methods for thermal energy research are still in the early stage, with applications spanning from atomistic materials discovery to system-level optimization.
Research Acceleration in Self‐Driving Labs: Technological Roadmap toward Accelerated Materials and Molecular Discovery
Self-driving labs integrating robotics, AI, and automated experimentation have the potential to accelerate materials and molecular discovery by 10–100 times, but are still in the technological roadmap phase.
Machine Learning and Optoelectronic Materials Discovery: A Growing Synergy
Machine learning techniques are speeding up ab initio calculations and guiding experimental synthesis of optoelectronic materials like perovskites and organic semiconductors.
Machine Learning Orchestrating the Materials Discovery and Performance Optimization of Redox Flow Battery
Integration of high-throughput computational screening and machine learning (including reinforcement learning, Bayesian optimization, and generative models) is accelerating materials discovery for redox flow batteries.