Artificial intelligence (AI) and human intelligence share deep thermodynamic foundations: both operate as systems that convert environmental energy and entropy into ordered, predictive structures under the constraints imposed by boundary conditions. Zentropy theory, an integration of quantum mechanics and statistical mechanics, provides a rigorous foundation for describing configurations and their contributions to free energy landscapes. Building on this foundation, the zentropy-enhanced neural network (ZENN) framework extends thermodynamic principles to complex data-driven systems by decomposing a system into independent configurations, each defined by its own internal energy and intrinsic entropy, which encode information across progressively finer scales down to quantum pure-state configurations, and integrating these contributions through a zentropy theory layer to provide a unified description of the system. This architecture not only improves predictive performance and stability across scientific benchmarks but also introduces two intrinsic mechanisms for AI safety: structural containment via configuration partitioning, and dynamic containment via zentropy-based regulation of driving forces and stability. By linking thermodynamic reasoning, statistics, and modern AI, ZENN offers a principled and physically grounded path toward safe, interpretable, and internally contained AI systems capable of scaling toward higher intelligence while remaining aligned with the laws that govern all viable systems and their evolutions. Zentropy theory thus represents an all-scale, pan-displinary framework with the configuration scale tailorable to desired granularity and precision.
ISSN: 3050-287X
AI for Science is an interdisciplinary and international peer-reviewed gold open access journal committed to publishing high-impact original research, reviews, and perspectives that highlight the transformative applications of artificial intelligence (AI) in driving scientific innovation. It is an affiliated journal of Dongguan Institute of Materials Science and Technology, CAS, China.
Free for readers. All article publication charges currently paid by the Dongguan Institute of Materials Science and Technology, Chinese Academy of Sciences.
- The following article is Open accessPerspectives on thermodynamics for AI and AI safety
Zi-Kui Liu and Bing Li 2026 AI Sci. 2 022001
- The following article is Open accessBeyond accuracy: comprehensive alignment for AI-driven molecular design
Qiang Zhang et al 2026 AI Sci. 2 023001
View article, Beyond accuracy: comprehensive alignment for AI-driven molecular designPDF, Beyond accuracy: comprehensive alignment for AI-driven molecular designArtificial intelligence (AI) is rapidly transforming biomolecular and material design, enabling advances in protein engineering, drug discovery, and material innovation. Yet, growing evidence shows that AI-generated candidates often violate physical laws, diverge from scientific objectives, or neglect safety and regulatory principles—revealing a fundamental misalignment between computational outputs and real-world requirements. Here, we propose comprehensive alignment, a framework that links AI systems not only to statistical data distributions but also to natural laws, scientific goals, and responsible research principles. Drawing examples across proteins, drugs, and materials, we illustrate how neglecting any alignment layer can yield unstable folds, unsynthesizable molecules, or unsafe outcomes. We then outline strategies to bridge these gaps through enriched datasets, hybrid physics-informed architectures, multi-objective evaluation, and closed-loop feedback. Together, these changes recast AI from a one-way generator into a trustworthy copilot—capable of accelerating discovery while remaining anchored in the principles that govern scientific validity and societal responsibility.
- The following article is Open accessLearning atomic representations for data-driven materials design
Zhenyao Fang et al 2026 AI Sci. 2 013001
View article, Learning atomic representations for data-driven materials designPDF, Learning atomic representations for data-driven materials designLearning the latent representations of atomic structures has become central to the application of machine learning (ML) in materials science, as such representations provide a unified framework for connecting atomic structures to material properties. Early physics-inspired descriptors facilitated efficient prediction of selected properties but were limited in flexibility and transferability. Recent advances in graph-based representations and graph neural networks (GNNs) have enabled data-driven feature learning frameworks that capture complex chemical environments, long-range interactions, and symmetry-governed responses directly from atomic structures. In this Perspective, we review recent progress in atomic representation learning for crystalline materials, with an emphasis on GNN architectures for predicting scalar, spectral, and tensorial properties. We discuss emerging challenges and opportunities related to high-fidelity datasets, model interpretability, and the integration of ML predictions with experimentally relevant phenomena, including disorder, dynamics, and finite-temperature effects. Finally, we outline future directions in which representation learning serves as a foundation for inverse materials design, leading to the systematic discovery and optimization of materials with targeted functional properties.
- The following article is Open accessSurface stability modeling with universal machine learning interatomic potentials: a comprehensive cleavage energy benchmarking study
Ardavan Mehdizadeh and Peter Schindler 2025 AI Sci. 1 025002
View article, Surface stability modeling with universal machine learning interatomic potentials: a comprehensive cleavage energy benchmarking studyPDF, Surface stability modeling with universal machine learning interatomic potentials: a comprehensive cleavage energy benchmarking studyMachine learning interatomic potentials (MLIPs) have revolutionized computational materials science by bridging the gap between quantum mechanical accuracy and classical simulation efficiency, enabling unprecedented exploration of materials properties across the periodic table. Despite their remarkable success in predicting bulk properties, no systematic evaluation has assessed how well these universal MLIPs (uMLIPs) can predict cleavage energies, a critical property governing fracture, catalysis, surface stability, and interfacial phenomena. Here, we present a comprehensive benchmark of 19 state-of-the-art uMLIPs for cleavage energy prediction using our previously established density functional theory database of 36 718 slab structures spanning elemental, binary, and ternary metallic compounds. We evaluate diverse architectural paradigms, analyzing their performance across chemical compositions, crystal systems, thickness, and surface orientations. Our zero-shot evaluation results reveal that training data composition dominates architectural sophistication: models trained on the Open Materials 2024 (OMat24) dataset, which emphasizes non-equilibrium configurations, achieve mean absolute percentage errors below 6% and correctly identify the thermodynamically most stable surface terminations in 87% of cases, without any explicit surface energy training. In contrast, architecturally identical models trained on equilibrium-only datasets show five-fold higher errors, while models trained on surface-adsorbate data fail catastrophically with a 17-fold degradation. Remarkably, simpler architectures trained on appropriate data achieve comparable accuracy to complex transformers while offering 10–100× computational speedup. These findings fundamentally reframe MLIP development priorities: highlighting that strategic training-data generation with appropriate non-equilibrium sampling deserves equal or greater attention than architectural complexity.
- The following article is Open accessQCell: comprehensive quantum-mechanical dataset spanning diverse biomolecular fragments
Adil Kabylda et al 2026 AI Sci. 2 025003
View article, QCell: comprehensive quantum-mechanical dataset spanning diverse biomolecular fragmentsPDF, QCell: comprehensive quantum-mechanical dataset spanning diverse biomolecular fragmentsRecent advances in machine learning force fields (MLFFs) are revolutionizing molecular simulations by bridging the gap between quantum-mechanical (QM) accuracy and the computational efficiency of mechanistic potentials. However, the development of reliable MLFFs for biomolecular systems remains constrained by the scarcity of high-quality, chemically diverse QM datasets that span all of the major classes of biomolecules expressed in living cells. Crucially, such a comprehensive dataset must be computed using non-empirical or minimally empirical approximations to solving the Schrödinger equation. To address these limitations, we introduce the QCell dataset – a curated collection of 525k new QM calculations for biomolecular fragments encompassing carbohydrates, nucleic acids, lipids, dimers, and ion clusters. QCell complements existing datasets, bringing the total number of available data points to 41 million molecular systems, all calculated using hybrid density functional theory with nonlocal many-body dispersion interactions, as captured by the PBE0+MBD(-NL) level of quantum mechanics. The QCell dataset therefore provides a valuable resource for training next-generation MLFFs capable of modeling the intricate interactions that govern biomolecular dynamics beyond small molecules and proteins.
- The following article is Open accessMOSES: combining automated ontology construction with a multi-agent system for explainable chemical knowledge reasoning
Yingkai Sun et al 2026 AI Sci. 2 015001
View article, MOSES: combining automated ontology construction with a multi-agent system for explainable chemical knowledge reasoningPDF, MOSES: combining automated ontology construction with a multi-agent system for explainable chemical knowledge reasoningThe vast and multiscale nature of chemical knowledge—from molecular structures to material properties—presents significant challenges for both human researchers and artificial intelligence (AI) systems. While large language models (LLMs) can process chemical information, they operate as black boxes without transparent reasoning. Here, we present our multi-agent ontology system for explainable knowledge synthesis (MOSES), a framework that combines automated knowledge organization with multi-agent collaboration to create an AI system for interpretable chemical knowledge reasoning. Using supramolecular chemistry as a testbed, we automatically constructed an ontology of over 10 000 classes from 52 publications and developed a multi-agent system that enables transparent knowledge retrieval and reasoning. Evaluations by human experts and LLMs show that MOSES significantly outperforms chemistry-oriented LLMs and leading general-purpose LLMs—including GPT-4.1 and o3—as well as GraphRAG-augmented GPT-4.1 models, on complex chemical questions, achieving superior scores in both direct assessments and Elo ratings. MOSES’s traceable reasoning paths reveal how it constructs answers through iterative refinement rather than probabilistic generation. However, we observe an asymmetry in handling positive versus negative knowledge claims, underscoring fundamental challenges in open-world reasoning. Our work demonstrates a pathway toward AI systems that can reason over complex scientific knowledge in a transparent and explainable manner.
- The following article is Open accessFrom prediction to discovery: AI as an observatory of physical organization in protein space
Yuxiang Zheng et al 2026 AI Sci. 2 032001
View article, From prediction to discovery: AI as an observatory of physical organization in protein spacePDF, From prediction to discovery: AI as an observatory of physical organization in protein spaceArtificial intelligence (AI) has transformed protein science from a data-sparse field into one increasingly rich in model-derived information. Predicted structures, sequence embeddings, confidence measures, mutation scores, inverse-design outputs, and generative ensembles provide complementary views of protein sequence, structure, dynamics, function, and designability. In this review, we develop the view of AI as an observatory of protein systems, shifting the question from how AI can be applied to protein problems to what successful models have learned about the constraints that shape proteins: physical constraints on structure and dynamics, statistical regularities in learned representations, and evolutionary constraints on sequence variation and design. This perspective is developed through three large-scale patterns exposed by AI-derived observables: the global structural landscape of the predicted protein Universe, proteome-scale relations between folding topology and native-state dynamics, and the organization of sequence, structure, and function into shared, searchable multimodal spaces. We then discuss how uncertainty analysis, perturbation and contrastive scoring, representation decomposition, physically informed probes, and experimental benchmarking extract interpretable information from these signals, and how the resulting descriptions connect to principles of folding, flexibility, evolutionary filtering, functional response, and design feasibility. At the same time, AI-derived observables are not direct physical measurements, but compressed, model-dependent readouts whose meaning requires systematic calibration. Thisperspective positions AI as both a predictive instrument and a systematic observational interface through which the organizational principles linking protein structure, dynamics, evolution, function, and design can be quantitatively probed and physically interpreted.
- The following article is Open accessTorchSim: an efficient atomistic simulation engine in PyTorch
Orion Cohen et al 2025 AI Sci. 1 025003
View article, TorchSim: an efficient atomistic simulation engine in PyTorchPDF, TorchSim: an efficient atomistic simulation engine in PyTorchWe introduce TorchSim, an open-source atomistic simulation engine tailored for the machine learned interatomic potential (MLIP) era. By rewriting core atomistic simulation primitives in PyTorch, TorchSim can achieve orders of magnitude acceleration for popular MLIPs. Unlike existing molecular dynamics (MD) packages, which simulate one system at a time, TorchSim performs batched simulations that efficiently utilize modern GPUs by evolving multiple systems concurrently. TorchSim supports MD integrators, structural relaxation optimizers, both machine-learned and classical interatomic potentials (such as Lennard–Jones, Morse, soft-sphere), batching with automatic memory management, differentiable simulation, and integration with popular materials informatics tools.
- The following article is Open accessBioinspired123D: generative 3D modeling system for bioinspired structures
Rachel K. Luu and Markus J. Buehler 2026 AI Sci. 2 025004
View article, Bioinspired123D: generative 3D modeling system for bioinspired structuresPDF, Bioinspired123D: generative 3D modeling system for bioinspired structuresGenerative AI has made rapid progress in text, image, and video synthesis, yet text-to-3D modeling for scientific design remains particularly challenging due to limited controllability and high computational cost. Most existing 3D generative methods rely on meshes, voxels, or point clouds which can be costly to train and difficult to control. We introduce Bioinspired123D, a lightweight and modular code-as-geometry pipeline that generates fabricable 3D structures directly through parametric programs rather than dense visual representations. At the core of Bioinspired123D is Bioinspired3D, a compact language model finetuned to translate natural language design cues into Blender Python scripts encoding smooth, biologically inspired geometries. We curate a domain-specific dataset of over 4000 bioinspired and geometric design scripts spanning helical, cellular, and tubular motifs with parametric variability. The dataset is expanded and validated through an automated large language model-driven, blender-based quality control pipeline. Bioinspired3D is then embedded in a graph-based agentic framework that integrates multimodal retrieval-augmented generation and a vision-language model critic to iteratively evaluate, critique, and repair generated scripts. We evaluate performance on a new benchmark for 3D geometry script generation and show that Bioinspired123D demonstrates a near fourfold improvement over its non-finetuned base model, while also outperforming substantially larger state-of-the-art language models despite using far fewer parameters and compute. By prioritizing code-as-geometry representations, Bioinspired123D enables compute-efficient, controllable, and interpretable text-to-3D generation, lowering barriers to AI driven scientific discovery in materials and structural design.
- The following article is Open accessAI-assisted wafer-scale exfoliation and transfer of 2D materials: status, challenges and perspectives
Haoyu Ge et al 2025 AI Sci. 1 013002
View article, AI-assisted wafer-scale exfoliation and transfer of 2D materials: status, challenges and perspectivesPDF, AI-assisted wafer-scale exfoliation and transfer of 2D materials: status, challenges and perspectivesMoving two-dimensional (2D) materials from lab to industry requires breakthroughs in scalable exfoliation and transfer methods. While traditional mechanical exfoliation methods can produce high-quality flakes, they suffer from poor reproducibility and low yield. In recent years, metal-assisted exfoliation techniques have significantly improved monolayer yield and structural uniformity. Furthermore, scalable transfer strategies such as polyvinyl alcohol-assisted transfer and van der Waals integration have achieved cleaner interfaces and higher alignment accuracy. However, manual operation remains a major limitation to consistency and efficiency. Artificial intelligence (AI) is emerging as a transformative tool, enabling intelligent control of the exfoliation and transfer process through real-time parameter optimization, crack prevention, and path planning. Deep learning architectures facilitate layer identification and defect detection, while reinforcement learning enables high-precision autonomous robotic manipulation. This article systematically reviews the latest advances in the field of 2D material exfoliation and transfer, highlighting the important role of AI in addressing core process bottlenecks and enabling the scalable, reliable, and automated fabrication of 2D materials.
- The following article is Open accessFrom prediction to discovery: AI as an observatory of physical organization in protein space
Yuxiang Zheng et al 2026 AI Sci. 2 032001
View article, From prediction to discovery: AI as an observatory of physical organization in protein spacePDF, From prediction to discovery: AI as an observatory of physical organization in protein spaceArtificial intelligence (AI) has transformed protein science from a data-sparse field into one increasingly rich in model-derived information. Predicted structures, sequence embeddings, confidence measures, mutation scores, inverse-design outputs, and generative ensembles provide complementary views of protein sequence, structure, dynamics, function, and designability. In this review, we develop the view of AI as an observatory of protein systems, shifting the question from how AI can be applied to protein problems to what successful models have learned about the constraints that shape proteins: physical constraints on structure and dynamics, statistical regularities in learned representations, and evolutionary constraints on sequence variation and design. This perspective is developed through three large-scale patterns exposed by AI-derived observables: the global structural landscape of the predicted protein Universe, proteome-scale relations between folding topology and native-state dynamics, and the organization of sequence, structure, and function into shared, searchable multimodal spaces. We then discuss how uncertainty analysis, perturbation and contrastive scoring, representation decomposition, physically informed probes, and experimental benchmarking extract interpretable information from these signals, and how the resulting descriptions connect to principles of folding, flexibility, evolutionary filtering, functional response, and design feasibility. At the same time, AI-derived observables are not direct physical measurements, but compressed, model-dependent readouts whose meaning requires systematic calibration. Thisperspective positions AI as both a predictive instrument and a systematic observational interface through which the organizational principles linking protein structure, dynamics, evolution, function, and design can be quantitatively probed and physically interpreted.
- The following article is Open accessPerspectives on thermodynamics for AI and AI safety
Zi-Kui Liu and Bing Li 2026 AI Sci. 2 022001
View article, Perspectives on thermodynamics for AI and AI safetyPDF, Perspectives on thermodynamics for AI and AI safetyArtificial intelligence (AI) and human intelligence share deep thermodynamic foundations: both operate as systems that convert environmental energy and entropy into ordered, predictive structures under the constraints imposed by boundary conditions. Zentropy theory, an integration of quantum mechanics and statistical mechanics, provides a rigorous foundation for describing configurations and their contributions to free energy landscapes. Building on this foundation, the zentropy-enhanced neural network (ZENN) framework extends thermodynamic principles to complex data-driven systems by decomposing a system into independent configurations, each defined by its own internal energy and intrinsic entropy, which encode information across progressively finer scales down to quantum pure-state configurations, and integrating these contributions through a zentropy theory layer to provide a unified description of the system. This architecture not only improves predictive performance and stability across scientific benchmarks but also introduces two intrinsic mechanisms for AI safety: structural containment via configuration partitioning, and dynamic containment via zentropy-based regulation of driving forces and stability. By linking thermodynamic reasoning, statistics, and modern AI, ZENN offers a principled and physically grounded path toward safe, interpretable, and internally contained AI systems capable of scaling toward higher intelligence while remaining aligned with the laws that govern all viable systems and their evolutions. Zentropy theory thus represents an all-scale, pan-displinary framework with the configuration scale tailorable to desired granularity and precision.
- The following article is Open accessAI-guided design of optoelectronic molecularly modified CH3NH3PbI3 perovskite thin films with improved aqueous stability
Lei Zhang et al 2026 AI Sci. 2 025006
View article, AI-guided design of optoelectronic molecularly modified CH3NH3PbI3 perovskite thin films with improved aqueous stabilityPDF, AI-guided design of optoelectronic molecularly modified CH3NH3PbI3 perovskite thin films with improved aqueous stabilityThe inherent vulnerability of halide perovskite films to moisture and water exposure severely restricts their stability, posing a challenge for industrial implementation. In this study, photoelectrochemical experiments, machine learning, and first-principles calculations are employed to accelerate the design toward aqueous-stable lead halide perovskite thin-film materials. A molecularly modified halide perovskite dataset incorporating diverse design parameters is constructed, enabling the development of a machine learning model that predicts a complex molecularly modified perovskite system that offers decent photocurrent in aqueous-based hostile environments. Specifically, a dye-modified MAPbI3 material, with ethyl red deposited as a molecular modifier on top of the perovskite thin film and an equimolar precursor ratio (PbI2: MAI = 1:1), is subsequently verified experimentally. The resulting CH3NH3PbI3 film achieves an improved photocurrent in aqueous solution and an enhanced photogenerated current retention rate of 98.56% in water after 400 s. Density functional theory reveals the atomic-scale origins underlying the machine-learning-predicted system, including the intimate interfacial contact between the molecular adsorbate and the perovskite substrate, as well as the resulting optimal optoelectronic properties. This study underscores the critical role of molecular composition and processing conditions in enhancing the aqueous stability of halide perovskites and demonstrates an accurate data-driven framework that enables AI-accelerated stability prediction and materials design.
- The following article is Open accessD3REAM: a framework for full-space inverse materials design with target properties via machine learning models and global optimization algorithms
Guanjian Cheng et al 2026 AI Sci. 2 025005
View article, D3REAM: a framework for full-space inverse materials design with target properties via machine learning models and global optimization algorithmsPDF, D3REAM: a framework for full-space inverse materials design with target properties via machine learning models and global optimization algorithmsFull-space inverse materials design (FSIMD) aims to identify materials with target properties by exploring the complete combinatorial space of possible elements, compositions, and structures, yet efficiently navigating the vast and high-dimensional materials space remains a fundamental challenge. Here, we present D3REAM (Data-Driven Design and Rapid Exploration for Advanced Materials), an open-source framework that unifies machine learning (ML) models and global optimization algorithms (OAs) for FSIMD with target properties. D3REAM integrates diverse universal ML interatomic potentials and universal property prediction models with multiple optimization strategies, including Bayesian optimization, swarm intelligence, and multi-objective OAs, to efficiently search for structures that meet specified target properties. Within D3REAM, a crystal structure is represented by three fundamental descriptors, i.e. atomic type (A), composition (C), and structural configuration (S). By adjusting the search domains of the (A, C, S) space, D3REAM can flexibly perform crystal structure prediction, variable-composition inverse materials design, and FSIMD. The framework combines ML with physics-informed search, and incorporates adaptive search-space pruning to construct a unified and efficient platform for accelerating materials discovery and optimization. Moreover, D3REAM is implemented in a modular and extensible architecture, enabling seamless integration with external simulation tools and OAs. In this paper, we focus on descriptions of the implementation and applications of the D3REAM framework.
- The following article is Open accessBioinspired123D: generative 3D modeling system for bioinspired structures
Rachel K. Luu and Markus J. Buehler 2026 AI Sci. 2 025004
View article, Bioinspired123D: generative 3D modeling system for bioinspired structuresPDF, Bioinspired123D: generative 3D modeling system for bioinspired structuresGenerative AI has made rapid progress in text, image, and video synthesis, yet text-to-3D modeling for scientific design remains particularly challenging due to limited controllability and high computational cost. Most existing 3D generative methods rely on meshes, voxels, or point clouds which can be costly to train and difficult to control. We introduce Bioinspired123D, a lightweight and modular code-as-geometry pipeline that generates fabricable 3D structures directly through parametric programs rather than dense visual representations. At the core of Bioinspired123D is Bioinspired3D, a compact language model finetuned to translate natural language design cues into Blender Python scripts encoding smooth, biologically inspired geometries. We curate a domain-specific dataset of over 4000 bioinspired and geometric design scripts spanning helical, cellular, and tubular motifs with parametric variability. The dataset is expanded and validated through an automated large language model-driven, blender-based quality control pipeline. Bioinspired3D is then embedded in a graph-based agentic framework that integrates multimodal retrieval-augmented generation and a vision-language model critic to iteratively evaluate, critique, and repair generated scripts. We evaluate performance on a new benchmark for 3D geometry script generation and show that Bioinspired123D demonstrates a near fourfold improvement over its non-finetuned base model, while also outperforming substantially larger state-of-the-art language models despite using far fewer parameters and compute. By prioritizing code-as-geometry representations, Bioinspired123D enables compute-efficient, controllable, and interpretable text-to-3D generation, lowering barriers to AI driven scientific discovery in materials and structural design.
- The following article is Open accessFrom prediction to discovery: AI as an observatory of physical organization in protein space
Yuxiang Zheng et al 2026 AI Sci. 2 032001
View article, From prediction to discovery: AI as an observatory of physical organization in protein spacePDF, From prediction to discovery: AI as an observatory of physical organization in protein spaceArtificial intelligence (AI) has transformed protein science from a data-sparse field into one increasingly rich in model-derived information. Predicted structures, sequence embeddings, confidence measures, mutation scores, inverse-design outputs, and generative ensembles provide complementary views of protein sequence, structure, dynamics, function, and designability. In this review, we develop the view of AI as an observatory of protein systems, shifting the question from how AI can be applied to protein problems to what successful models have learned about the constraints that shape proteins: physical constraints on structure and dynamics, statistical regularities in learned representations, and evolutionary constraints on sequence variation and design. This perspective is developed through three large-scale patterns exposed by AI-derived observables: the global structural landscape of the predicted protein Universe, proteome-scale relations between folding topology and native-state dynamics, and the organization of sequence, structure, and function into shared, searchable multimodal spaces. We then discuss how uncertainty analysis, perturbation and contrastive scoring, representation decomposition, physically informed probes, and experimental benchmarking extract interpretable information from these signals, and how the resulting descriptions connect to principles of folding, flexibility, evolutionary filtering, functional response, and design feasibility. At the same time, AI-derived observables are not direct physical measurements, but compressed, model-dependent readouts whose meaning requires systematic calibration. Thisperspective positions AI as both a predictive instrument and a systematic observational interface through which the organizational principles linking protein structure, dynamics, evolution, function, and design can be quantitatively probed and physically interpreted.
- The following article is Open accessPerspectives on thermodynamics for AI and AI safety
Zi-Kui Liu and Bing Li 2026 AI Sci. 2 022001
View article, Perspectives on thermodynamics for AI and AI safetyPDF, Perspectives on thermodynamics for AI and AI safetyArtificial intelligence (AI) and human intelligence share deep thermodynamic foundations: both operate as systems that convert environmental energy and entropy into ordered, predictive structures under the constraints imposed by boundary conditions. Zentropy theory, an integration of quantum mechanics and statistical mechanics, provides a rigorous foundation for describing configurations and their contributions to free energy landscapes. Building on this foundation, the zentropy-enhanced neural network (ZENN) framework extends thermodynamic principles to complex data-driven systems by decomposing a system into independent configurations, each defined by its own internal energy and intrinsic entropy, which encode information across progressively finer scales down to quantum pure-state configurations, and integrating these contributions through a zentropy theory layer to provide a unified description of the system. This architecture not only improves predictive performance and stability across scientific benchmarks but also introduces two intrinsic mechanisms for AI safety: structural containment via configuration partitioning, and dynamic containment via zentropy-based regulation of driving forces and stability. By linking thermodynamic reasoning, statistics, and modern AI, ZENN offers a principled and physically grounded path toward safe, interpretable, and internally contained AI systems capable of scaling toward higher intelligence while remaining aligned with the laws that govern all viable systems and their evolutions. Zentropy theory thus represents an all-scale, pan-displinary framework with the configuration scale tailorable to desired granularity and precision.
- The following article is Open accessTowards intelligent design of metallic glasses: a data-driven pathway for closing the theory-experiment loop
Huanrong Liu et al 2026 AI Sci. 2 012001
View article, Towards intelligent design of metallic glasses: a data-driven pathway for closing the theory-experiment loopPDF, Towards intelligent design of metallic glasses: a data-driven pathway for closing the theory-experiment loopAs advanced amorphous materials with superior mechanical, physical, and chemical properties, metallic glasses (MGs) hold significant promise for a wide range of applications. However, their rational design and precise property control have long been impeded by notable challenges. These obstacles largely arise from the inherent compositional and structural complexity of MGs, which not only slows empirical trial-and-error experimentation but also limits the scalability of computationally intensive first-principles simulations. In recent years, machine learning has emerged as a transformative tool, offering unprecedented capabilities to decode these intricate relationships and overcome conventional research limitations. Here we provide a systematic overview of machine-learning-guided investigations of MGs and their associated data pipelines, centering on two key paradigms: the feature-driven ‘Keplerian’ data pipeline and a next-generation theory-experiment-aligned pipeline designed to close the gap between simulation and experiment. By breaking down both paradigms into modular workflow components, we underscore the essential roles of data standardization, interpretable feature engineering, and physics-informed validation in constructing a reliable research framework. We anticipate that a sufficiently robust, efficient, and generalizable data pipeline will not only unlock novel scientific insights and accelerate material discovery but also propel the field from a ‘trial-and-error’ approach toward an era of intelligent and principled design.
- The following article is Open accessArtificial intelligence for fibrous network design and mechanics
Yunhao Yang et al 2025 AI Sci. 1 012001
View article, Artificial intelligence for fibrous network design and mechanicsPDF, Artificial intelligence for fibrous network design and mechanicsFibrous networks are critical structural motifs underpinning numerous biological and engineering materials. Their complex mechanics, governed by fiber properties and topological architecture, therefore require advanced modeling and optimization strategies. This review presents a comprehensive synthesis of recent advances in artificial intelligence (AI)–assisted design, characterization, and optimization of fibrous networks. We explore how deep generative models enable the creation of ordered and disordered architectures with tailored properties, how machine learning facilitates structure–property prediction across multiple physical fields and spatial dimensions, and how reinforcement learning accelerates performance-driven topological optimization. Emphasis is placed on the integration of multi-scale data, physics-informed learning, and explainable AI to enhance design fidelity and interpretability. We conclude by outlining future opportunities for autonomous material systems, including closed-loop discovery platforms and multi-physics integration, positioning AI as a transformative force in fibrous materials innovation.
Trending on Altmetric
Journal links
Journal information
- 2025-present
AI for Science
Online ISSN: 3050-287X