AI-driven materials design: a mini-review

TL;DR

This mini-review surveys AI-enabled materials design, contrasting traditional forward screening with inverse design and detailing how advances in adaptive methods and deep generative models have reshaped the discovery process. It covers forward screening, evolutionary algorithms, adaptive/instrumented approaches, and state-of-the-art generative frameworks (VAEs, GANs, diffusion models, and LLMs), highlighting how conditional generation enables property-targeted material design. The article discusses key challenges—thermodynamic stability, experimental feasibility, data quality, and extrapolation beyond known spaces—and outlines a roadmap for robust, autonomous closed-loop discovery combining computation and autonomous experimentation. The work provides a concise guide to techniques, progress, and future directions for rapid, targeted discovery of functional materials with high technological relevance.

Abstract

Materials design is an important component of modern science and technology, yet traditional approaches rely heavily on trial-and-error and can be inefficient. Computational techniques, enhanced by modern artificial intelligence (AI), have greatly accelerated the design of new materials. Among these approaches, inverse design has shown great promise in designing materials that meet specific property requirements. In this mini-review, we summarize key computational advancements for materials design over the past few decades. We follow the evolution of relevant materials design techniques, from high-throughput forward machine learning (ML) methods and evolutionary algorithms, to advanced AI strategies like reinforcement learning (RL) and deep generative models. We highlight the paradigm shift from conventional screening approaches to inverse generation driven by deep generative models. Finally, we discuss current challenges and future perspectives of materials inverse design. This review may serve as a brief guide to the approaches, progress, and outlook of designing future functional materials with technological relevance.

Figures (7)

• Figure 1: Evolutionary trend of materials design paradigm for forward vs. inverse design since 2014, by number of publications. The publications presented in this figure are retrieved from the Web of Science database with keywords related to automatic materials design and discovery. Publications with inverse design are filtered based on additional keywords with various AI-based inverse-design techniques. Milestone AI techniques and frameworks are labeled on the top panel along the time line, with deep generative models marked bold in brown.
• Figure 2: Material discovery based on forward screening.(a) Schematics of forward screening, which filters material candidates by high-throughput calculation of target properties. Such calculations can be accelerated by ML surrogate models. (b) Crystal graph convolutional neural networks (CGCNN) as a prominent framework of GNN for materials property prediction and screeningxie2018crystal. (c) A five-stage high-throughput computational screening process (purple disks) for optoelectronic semiconductorsluo2021high. Candidates are filtered based on stability and descriptor-based refinements. Machine learning speeds up predictions while balancing computational cost (red arrows) and accuracy (green arrows). (d) The screening workflow for identifying 2D ferromagnetic (FM) materialskabiraj2020high. Spin configurations undergo energy calculations to classify materials as FM or antiferromagnetic (AFM). FM candidates are further refined through Hamiltonian fitting and Monte Carlo simulations to estimate the Curie temperature. Reproduced with permissionxie2018crystalluo2021highkabiraj2020high.
• Figure 3: Evolutionary algorithms (EAs) for materials inverse design. The top diagrams show the schematics of (a) genetic algorithm (GA), where candidate materials undergo mutation (A to D), crossover (B & C exchanged), and selection driven by a target function calculator; (b) Particle swarm optimization (PSO), where candidate solutions evolve by following individual (purple arrows) and the global optimum position (star) in a search space; and (c) Monte Carlo tree search (MCTS), where a decision tree is expanded (green solid arrow) iteratively using simulations (blue dashed circle) and back-propagation (brown dashed arrow) to identify optimal candidates. (d) Evolutionary routes that explore the candidate structures of tetracene in GAtom2023inverse. Different evolutionary strategies, including crossover, frame rotation, and angle strain, guide structural transformations between space groups, and various selection criteria (Energy-based, Niching, and SF+Energy) influence the optimization pathway. (e) The evolutionary process of the stable structure for Pt–Pd–Au nanoparticles with 3285 atoms in PSOfan2015structural, with corresponding nanoparticle structures shown at a three different evolutionary stages. (f) Comparison of energy convergence between MCTS (blue curve) and random sampling (orange curve) of a four-atom silver cluster (green spheres). MCTS reaches faster convergence and identifies an optimal structure with lower-energy (label "A") compared to random sampling (label "B")banik2023continuous. Reproduced with permissiontom2023inversefan2015structuralbanik2023continuous.
• Figure 4: Adaptive and interactive approaches for inverse design. (a) A schematic to illustrate the adaptive approach, where the inverse design model is associated with the guidance and feedback/update loops to the experiments and computation. (b) Graphical illustration of the BO used for chemical reaction shields2021bayesian. A Gaussian process surrogate model is fitted to data from an unknown objective (black dashed line), with the posterior mean (blue curve) shown with a shaded $2\sigma$ region. The expected utility curve (orange curve) guides the next experiment, which updates the current knowledge of the surrogate model. (c) Methodology illustration to design optimization of graded metamaterials for energy harvestingrosafalco2023reinforcement. The design process is framed as a Markov Decision Process, where physical constraints and interpolation points guide sequential decisions, and is iteratively optimized through reinforcement learning. (d) The illustration of autonomous materials discovery with the A-Lab szymanski2023autonomous. A-Lab integrates computation (top left), text mining (top middle), robotic synthesis (right), recipe optimization (bottom left), and phase identification (bottom right) to optimize reaction pathways. Machine learning suggests synthesis recipes, experimental robots execute synthesis and characterization, and diffraction analysis confirms phase identification. Reproduced with permissionshields2021bayesianrosafalco2023reinforcementszymanski2023autonomous.
• Figure 5: Variational autoencoder (VAE) approaches for inverse design.(a) The general idea behind the autoencoder (AE) and VAE approach: In an AE, the encoder maps inputs to deterministic points in latent space $Z$, while in a VAE, the latent space $Z$ is constrained by a probabilistic prior (e.g., Gaussian distribution). Compared to AE, the structured latent space of VAE enhances semantic continuity and can be used for data generation. (b) A common approach for inverse design with VAE via latent space search and optimization. The latent representation of each input structure can be trained with the corresponding properties by another small machine learning model to create a property map, in which the search and optimization algorithm can provide candidate latent vectors, from which the decoder can reconstruct them to obtain structures with the desired propertiesgomez2018automatic. (c) Jointly trained VAE between structure-property learningfallani2024inverse. The model aligns latent spaces ($z$) of molecular structures ($x$) and properties ($y$), enabling direct inverse mapping from properties (bottom left) to target structures (right) for efficient materials design. Reproduced with permissiongomez2018automaticfallani2024inverse.