Achieving global carbon neutrality requires a rapid transition toward a sustainable energy economy, a shift that fundamentally relies on the discovery of high-performance materials 1. Central to this transition are modern energy systems requiring inorganic crystals with tailored properties, such as solid-state batteries for safe storage 2-4, perovskite solar cells for efficient power generation 5-7, and catalysts that facilitate green hydrogen production 8-10. However, the discovery of such materials is hampered by the need to reconcile inherently conflicting properties within a vast chemical space estimated to exceed 10100 possible inorganic compounds 11. The sheer magnitude of this chemical landscape renders conventional trial-and-error methodologies prohibitively slow and inefficient. As traditional experimental approaches struggle to achieve the transformative breakthroughs required by modern energy demands 12, researchers have reached a critical juncture that necessitates innovative paradigms to explore this vast configurational space 13. Over the past decade, the discovery landscape has been defined by high-throughput (HT) screening 14-16. Under this “forward-design” framework, candidates are drawn from databases of documented crystal structures, with computational models employed to screen for specific performance metrics 17-19. However, this strategy is inherently retrospective; it operates more as a sophisticated filter than a creative source of discovery. Notably, it struggles to identify unconventional structures that deviate from existing structural archetypes. As a result, the identification of fundamentally new material classes remains elusive because the process is limited by the biases of human intuition and established chemical domains. The period between 2024 and 2025 has witnessed a fundamental paradigm shift in materials design, evolving from purely property-predictive models toward generative frameworks capable of creating entirely new structures from the ground up 20-22. Pioneering tools, such as MatterGen, have redefined the role of artificial intelligence (AI) in the field 23. These systems have shifted from the passive evaluation of candidate structures toward the autonomous generative design of new crystalline configurations. By leveraging sophisticated mathematical architectures and massive datasets, they have expanded the library of stable crystals by millions of entries. This transition from library-based filtering to goal-oriented inverse design holds immense potential to bridge the gap between computational theory and the experimental realization of next-generation clean energy materials. The discovery of energy materials is undergoing a fundamental shift from HT screening to generative inverse design. Although early architectures such as variational autoencoders and generative adversarial networks demonstrated potential, they often struggled with thermodynamic stability and complex constraints. The emergence of diffusion models has addressed these limitations by providing a robust framework for generating physically realistic inorganic structures through iterative denoising. Building on this diffusion-based paradigm, MatterGen represents a landmark by treating crystal structure generation as a joint diffusion process. As shown in Figure 1A, the model iteratively denoises random distributions of atom types, coordinates, and lattices into stable crystalline structures. The practical power of this architecture lies in its “on-demand” controllability. By incorporating adapter modules for fine-tuning 25 (Figure 1B), the model can be precisely steered toward specific chemical systems, symmetries, and target physical properties (Figure 1C). Benchmark comparisons (Figure 1D,E) confirm that MatterGen more than doubles the success rate in generating stable, unique, and new (SUN) materials compared to previous models such as crystal diffusion variational autoencoder (CDVAE) 26 and DiffCSP 27, whereas producing structures that are significantly closer to the local energy minima. (A–C) MatterGen workflow. An equivariant score network iteratively denoises structures (A, X, L) for property-conditioned (c) generation. Pretrained on stable materials and fine-tuned via adapters, the model enables the targeted design of novel materials with specific chemical or physical attributes (m) 23. (D, E) Performance metrics showing the percentage of SUN (stable, unique, novel) structures and the average root mean squared displacement (RMSD) between initial and DFT-relaxed structures 23. (F, G) Constrained generation methodology. A diffusion denoising framework utilizes geometric patterns—specifically Archimedean lattice templates—as structural initialization constraints, governed by lattice symmetry, atomic species, and spatial parameters 24. (A–E) Reprinted from Ref. 23. (F, G) Reprinted from Ref. 24. Whereas property-driven generation satisfies macroscopic targets, the discovery of quantum-energy materials often requires precise control over microscopic geometric patterns. To address this challenge, structural constraint integration in the GENerative model (SCIGEN) introduces a framework that incorporates structural priors without the need for extensive retraining. As shown in Figure 1F, the process begins with the definition of structural constraints, such as Archimedean or Lieb-like lattices, in which parameters such as atom types, bond lengths, and unit cell occupancy are initialized. The core innovation lies in the integration of these priors into the generative pathway. As illustrated in Figure 1G, the algorithm strategically applies a masking operator during each denoising step. This operator merges the diffused constrained structural elements () with the unconstrained segments (), effectively guiding the model to preserve specific topological features while allowing it to freely optimize the remaining atomic environment. This methodology ensures that the generated candidates satisfy the stringent geometric requirements essential for frustrated magnets and topological insulators while maintaining the high thermodynamic stability characteristic of diffusion-based generative models. The synergy between generative AI and high-throughput (HT) hardware has established a new paradigm for bridging the “synthesis gap”. By shifting from traditional trial-and-error to closed-loop validation, researchers can rapidly verify AI-predicted blueprints within complex device architectures. The iterative inverse design workflow for hole-transport materials in perovskite solar cells is shown in Figure 2A–D 5. This process was initiated with density functional theory (DFT)-derived descriptors for chemical mapping (Figure 2A), followed by validation using an HT synthesis and device characterization pipeline (Figure 2B,C). The experimental data were then fed back into the model to predict and identify optimal candidates, completing the closed-loop cycle (Figure 2D). The performance statistics in Figure 2E show that this approach led to a steady increase in power conversion efficiency (PCE). Personalized device optimization based on specific molecular properties achieved a peak certified PCE of 26.2% (stabilized at 25.9%), significantly outperforming standardized conditions (Figure 2F). (A) Generation of molecular descriptors via DFT calculations 5. (B) High-throughput (HTP) synthesis, purification, and characterization using an in-house platform 5. (C) Device integration and evaluation of synthesized molecules as hole-transporting materials (HTMs) in perovskite solar cells (PSCs) 5. (D) Iterative ML workflow: model training, performance prediction, and experimental feedback for closed-loop optimization 5. (E) Evolution of device PCE across iterations, normalized to poly(triarylamine) (PTAA)-based reference devices 5. (F) Comparison of device performance under standardized (blue) and molecular-property-based personalized (orange) optimization conditions 5 (Voc is open-circuit voltage, Jsc is short-circuit current density, FF is fill factor, PCE is power conversion efficiency). (G) Autonomous materials discovery enabled by the A-Lab platform 28. (H) Outcomes from targeted syntheses of DFT-predicted materials 28. (A–F) Reprinted with permission from J. Wu, et al., Science, 386, 1256–1264 (2024) 5. Copyright 2024 by the American Physical Society. (G, H) Reprinted from Ref. 28. Moving toward a generalized inorganic material discovery, Figure 2G illustrates the architecture of A-Lab, a fully automated platform that integrates literature-mined recipes with robotic powder dosing, heating, and X-ray diffraction characterization 28. By utilizing an active-learning algorithm to identify reaction pathways with the maximal driving force, the system autonomously refines its synthesis strategy. The efficiency of this platform is depicted in Figure 2H, where 36 out of 57 targets were successfully synthesized within a 17-day autonomous campaign. These results demonstrate that the integration of generative models with autonomous hardware can effectively compress the discovery-to-verification cycle from years to weeks. Although the achievements of generative AI and autonomous laboratories are transformative, the field is currently undergoing a rigorous “correction phase”. Recent critiques have sparked intense debate regarding the physical validity and true novelty of AI-driven discoveries, revealing challenges categorized as “AI hallucinations” and “characterization pitfalls”. A primary concern is whether generative models are truly navigating “unexplored” chemical space or merely repackaging known phases. As shown in Figure 3A–F, Juelsholt 29 reexamined the Ta–Cr–O system, specifically the compound TaCr2O6 claimed as a novel discovery by MatterGen. Crystallographic analysis suggests that the predicted structure is essentially a symmetry-transformed version of Ta1/2Cr1/2O2 (Figure 3G), a disordered rutile-type phase reported as early as 1972 and present in training datasets. Furthermore, the Rietveld refinement in Figure 3H demonstrates that the “novel” experimental pattern is indistinguishable from the existing disordered model. This “hallucination of novelty” underscores the risk of AI misinterpreting cation disorders as new ordered phases, highlighting the need for more rigorous deduplication against established databases, such as the Inorganic Crystal Structure Database. Crystal structures and powder x-ray diffraction (PXRD) analysis of Cr–Ta and Mg–Ni oxides. Crystal structures of Cr2TaO6 (A, B), Cr2/3Ta1/3O2 (C, D), and Ta1/2Cr1/2O2 (E, F) viewed along the a and c axes 29. (G) Symmetry operation transforming Ta1/2Cr1/2O2 into Ta1/3Cr2/3O2. Atom colors: Cr (blue), Ta (gold), and O (red) 29. (H) Rietveld refinement of the PXRD pattern for Ta1/2Cr1/2O2 29. (I) The distribution of errors in the 36 claimed “successful” syntheses. The “X” symbol denotes that the error is present: (1) poor fit, such that the fitted model is meaningless; (2) crystallographic information file (CIF) mismatch; (3) lack of cation ordering evidence; (4) previously reported compounds 30. (J–L) Comparison of structures and simulated/measured PXRD patterns for Mg3NiO4 (Materials Project and Szymanski et al. 28) and Mg2Ni2O4 (ICSD). Atom colors: Mg (orange), Ni (gray), and O (red) 30. (A–H) Reprinted from Ref. 29. (I–L) Reprinted from Ref. 28, 30. The reliability of unsupervised experimental validation has also encountered academic skepticism. Leeman et al. 30 identified four systematic error types across the “successful” syntheses reported by A-Lab, ranging from meaningless fits (Error 1) to the misidentification of already reported compounds (Error 4), as shown in Figure 3I. A representative case is the attempted synthesis of Mg3NiO4. Although the Materials Project predicted an ordered structure (Figure 3J), the experimental powder pattern provided by A-Lab (Figure 3K) did not match this prediction. Instead, the measured data closely resembled the simulated pattern of the known disordered MgNiO2 phase (Figure 3L). These discrepancies suggest that current AI workflows struggle to account for cation occupancy and kinetic pathways, often leading to the “discovery” of materials that are either structurally mischaracterized or already known in their disordered forms. The controversies surrounding MatterGen and A-Lab emphasize that “scaling” alone cannot substitute for physical accuracy. The reliance on DFT-calculated ground states often neglects temperature-dependent entropy and disordered states, which are critical for functional energy materials. To advance this field, the integration of deeper physical descriptors and the adoption of standardized, high-fidelity validation protocols are essential. Addressing these “hallucinations” will be crucial for transitioning AI from a high-volume candidate generator to a precision tool for genuine material innovation. The paradigm shift from discriminative screening to generative inverse design is increasingly defined by the integration of synthesis awareness and latent chemical reasoning. To surmount the “AI hallucinations” and persistent “synthesis gap” identified in current frameworks, future discovery pipelines must transcend pure structural sampling by embedding experimental and semantic priors. The DiffSyn framework 31 represents a landmark application in this direction by co-generating crystal structures alongside their corresponding synthesis routes. By utilizing a generative diffusion process to simultaneously optimize precursors and reaction conditions, it ensures that generated blueprints are inherently tied to experimental accessibility. A critical extension of this methodology involves coupling these synthesis-aware trajectories with differentiable physics gradients. By embedding thermodynamic potentials directly into the diffusion framework, future iterations can ensure that every generated step not only targets a viable synthesis path but also converges toward a global energy minimum, effectively reconciling kinetic feasibility with thermodynamic stability. MatLLMSearch has demonstrated that large language models (LLMs) can function as “innate crystal generators” by leveraging their internalized understanding of atomic arrangements 32. This application signifies a shift toward knowledge-driven discovery, where LLMs handle complex high-dimensional constraints through semantic reasoning rather than exhaustive spatial searching. The potential extension of this paradigm lies in the development of “agentic materials science”. In this vision, LLM-based agents would not only propose structures from latent chemical logic but also autonomously orchestrate differentiable physics engines and A-Lab hardware to iteratively refine operational strategies based on real-time feedback. By grounding this transformative generative power in the rigorous foundations of physical science, we can resolve the disconnect between algorithmic prediction and experimental verification, ultimately accelerating the realization of materials vital for a carbon-neutral future. Shengxian Liu: methodology, conceptualization, writing – original draft. Kan Tang: methodology, conceptualization, data curation. Jian Zhou: visualization, validation, supervision, writing – review and editing. Zhimei Sun: funding acquisition, validation, visualization, supervision, writing – review and editing. This work was supported by the Advanced Materials-National Science and Technology Major Project (Grant No. 2025ZD0618802) and the National Natural Science Foundation of China (Grant No. 52332005). The authors declare no conflicts of interest. The data that support the findings of this study are available from the corresponding author upon reasonable request.
Liu et al. (Sun,) studied this question.