1 Introduction
The integration of artificial intelligence (AI) into scientific research — often termed “AI for Science” (AI4S) — has emerged as a transformative paradigm that is reshaping how scientific discovery is conducted across disciplines Rather than merely applying AI as a computational tool, this paradigm represents a fundamental shift in methodology: AI enables researchers to extract hidden patterns from complex data, predict material properties with unprecedented accuracy, design novel devices, and even uncover new physical principles.
The present survey report provides a systematic overview of AI4S research based on a diverse portfolio of recent publications spanning atomic, molecular, and optical (AMO) physics, condensed matter physics, materials physics, quantum information, and interdisciplinary AI applications [
1−
17], complemented by critical perspectives from the broader scientific community. By synthesizing technical advancements, methodological evolution, unresolved bottlenecks, and prospective trajectories, this work aims to provide a comprehensive and critical overview of the current AI4S landscape for physical and material sciences.
2 Major research themes in AI for science
Recent literature demonstrates a set of distinctive yet intrinsically interconnected research themes (Fig. 1), representing the mainstream development trends and application frontiers of contemporary AI4S in physical disciplines.
2.1 Machine learning for spectroscopic analysis
Machine learning has become a powerful auxiliary tool to compensate for inherent physical limitations in spectral analytical techniques. In 2024, Hao
et al. [
1] provided a comprehensive review of machine learning applications in laser-induced breakdown spectroscopy (LIBS). LIBS is a spectroscopic technique with great potential for online and in-situ detection, but its quantification results have traditionally been limited by complex matrix effects and signal fluctuations [
1]. Machine learning offers a solution by intelligently correlating complex spectral data with qualitative and quantitative composition through multivariate regression models [
1]. The review systematically covers two main aspects: (i) data pre-processing — including spectral selection, variable reconstruction, and denoising — and (ii) machine learning methods for improved quantification that reduce matrix effects and spectral fluctuations [
1−
3]. Critically, several unresolved challenges were identified, including training data limitations, the disconnect between physical principles and algorithms, low generalization ability, and massive data processing requirements [
1,
3]. More critically, the mismatch between empirical algorithmic correlation and rigorous physical principles reveals a fundamental disciplinary contradiction in AI4S research: the trade-off between data-driven predictive flexibility and physical interpretability and constraint.
2.2 Machine learning for materials science
Machine learning is driving a paradigm shift in materials science from empirical trial-and-error iteration to systematic data-driven discovery. In 2024, Chong
et al. [
2] provided a broad review of machine learning advances in materials science, marking what they describe as a “paradigm shift” from traditional trial-and-error approaches to data-driven discovery. The review covers the full machine learning workflow: from basic concepts and algorithms to descriptors, databases, and emerging trends [
2]. Notably, Chong
et al.’s review traces the evolution of the field from numerical analysis to feature engineering, then to representation learning, and finally to inverse design, a progression that reflects the increasing sophistication of AI applications in materials discovery. Complementing this general perspective, Li
et al. [
3] offered a critical review that mainly focuses on compositionally complex alloys (CCAs), a class of materials whose huge compositional search space and strong structure–property coupling pose unique challenges. While sharing the overarching data-driven philosophy, Li
et al. [
3] delve into domain-specific tasks such as phase formation prediction (amorphous, solid solution, intermetallic, and mixed phases), property prediction (saturation magnetic flux density, Curie temperature, coercivity, and hardness), and single- or multi-objective optimization via active learning and metaheuristic algorithms. They also place particular emphasis on feature engineering — using thermodynamic parameters (mixing entropy, enthalpy, atomic-size difference) and electronic descriptors — as well as on the critical issues of data scarcity and imbalanced datasets, which are especially acute in magnetic CCA research. Both reviews underscore the transformative potential of ML for accelerating materials discovery, yet Li
et al. [
3] offer a more granular account of how domain knowledge (e.g., the Slater–Pauling curve, glass-forming ability criteria) can be integrated into ML workflows to enhance predictive performance and interpretability. Together, these complementary reviews illustrate that while general ML frameworks are broadly applicable, successful adoption in specific material systems demands careful tailoring of features, algorithms, and experimental feedback loops to overcome the inherent complexities of the target system.
Several representative studies have verified the practical effectiveness of optimized AI material prediction frameworks. One particularly striking example of this progress is the work by Xu and co-workers [
4] in 2025, who developed a hierarchical neural network (HNN) AI model to predict superconducting critical temperatures (
Tc). The challenge was formidable: neural network approaches to materials science often face a contradiction between the large number of descriptors and the small size of available datasets [
4]. The HNN model resolved this by incorporating 909 universal descriptors for inorganic compounds and achieving a test
R2 score of 95.6%. Most impressively, the model predicted
Tc for 45 new high-entropy alloy superconductors with a mean absolute percent error below 6% compared to experimental data [
4].
Complementing this work, Jiang
et al. [
5] have addressed the challenge of small sample datasets in materials discovery by combining Random Forests prescreening with the Sure Independence Screening and Sparsifying Operator (SISSO) method for complex feature selection. This work illustrates a crucial methodological insight: when data is scarce, the choice of feature selection and dimensionality reduction techniques can be as important as the choice of predictive model itself.
2.3 Memristors and neuromorphic computing: The hardware-AI synergy
Memristor-based in-memory computing systems have emerged as core hardware carriers for next-generation energy-efficient artificial intelligence, thereby fundamentally breaking the von Neumann separation bottleneck between data storage and computation [
6,
7]. As the physical foundation of neuromorphic intelligence, memristors enable the integration of storage, calculation, and dynamic nonlinear response, realizing highly efficient hardware-algorithm synergy for low-power neuromorphic computing.
Fundamental materials and mechanisms. Several review articles document the rapid evolution of memristor materials. In 2024, Niu
et al. [
8] systematically summarized two-dimensional ferroelectric memristors, which show great promise due to their atomic-scale thickness, high carrier mobility, and mechanical flexibility. The presence of natural dipoles in these materials enables resistive switching through polarization reversal [
8]. Liu
et al. [
6] have also provided a comprehensive overview of halide perovskite memristors, highlighting their unique electrical and optical properties — including ion migration, charge trapping, and high charge mobility — that make them highly promising for memristor applications.
Device innovations. At the device level, Lee and co-workers [
9] demonstrated a self-aligned TiO
x-based three-dimensional vertical memristor for high-density synaptic arrays. The device achieved stable bipolar switching, endurance exceeding 10
4 cycles, and retention beyond 10
4 seconds [
9]. Most importantly, the device successfully emulated biological synaptic functions including potentiation, depression, spike-rate-dependent plasticity, and spike-timing-dependent plasticity [
9].
Reservoir computing. Reservoir computing (RC) has become a frontier neuromorphic computing paradigm suitable for time-series dynamic processing, requiring only readout layer training and offering extreme efficiency [
10]. In 2024, Chen
et al. [
11] reviewed emerging memristors and their applications in RC, noting that physical RC systems — which utilize the nonlinear dynamics and short-term memory properties of memristors — offer comparable performance to digital RC systems but with lower energy consumption and greater robustness. Recently, Jang
et al. [
10] provided a concrete demonstration using NbO
x-based memristors with coexisting short-term and long-term memory characteristics. Their device, configured as a physical reservoir with all 16 possible states (4-bit), combined with an artificial neural network readout layer, achieved a pattern recognition accuracy of 92.34% on the modified MNIST dataset [
10]. Notably, it was found that short-term memory and long-term memory in the device coexist by adjusting the intensity of pulse stimulation — an important finding with implications for the development of adaptive, multi-timescale neuromorphic systems.
Future AI computing paradigms. Looking towards the future, Wang
et al. [
7] recently offered a perspective on the challenges and opportunities for memristors in modern AI computing paradigms. They highlight that transformer-based AI models, with their attention mechanisms that dynamically generate weights and require global parameter computations, pose unprecedented demands on memristor performance — including exceptional endurance, low latency, energy efficiency, and device uniformity. These demands will dominate the iterative upgrading of next-generation memristor materials, structures, and integration technologies [
7].
2.4 AI for optical communication and information processing
AI has demonstrated unique advantages in solving nonlinear distortion and turbulence interference in optical communication systems. In 2024, Shi
et al. [
11] demonstrated the application of convolutional neural networks to optical communication, achieving high-resolution recognition of fractional orbital angular momentum (FOAM) modes. Using an improved EfficientNetV2-based CNN, they achieved a recognition resolution as high as 0.001 — the first time such resolution has been reported. Under strong atmospheric turbulence, the model maintained recognition accuracies of 99.12% and 92.24% for 0.1 and 0.01 resolution modes, respectively. This work has significant implications for free-space optical communication, where FOAM modes offer the potential for greatly increased channel capacity [
11].
2.5 AI for quantum information and computing
The intersection of AI and quantum physics represents one of the most dynamic frontiers in AI4S for solving intractable quantum system simulation and control problems, with recent advances spanning quantum circuit optimization, quantum secure communication, and machine learning for quantum systems.
Graph neural networks for quantum circuit prediction. In 2026, Liu
et al. [
12] recently employed graph neural networks (GNNs) for quantum circuit output prediction. GNNs are particularly well-suited for this task because quantum circuits can be naturally represented as graphs, with qubits as nodes and quantum gates as edges. Most notably, a “Direct Comparison” scheme that directly predicts the relative performance of two circuits was proposed, which significantly outperforms the indirect approach by an average of 36.2%. This work exemplifies how classical machine learning techniques can be adapted to address challenges in quantum computing — a field where the exponential growth of Hilbert space makes traditional simulation methods impractical.
AI-assisted quantum gate optimization. In 2026, Fu, Wang, and Xiong [
13] addressed a critical bottleneck in superconducting quantum computing: fidelity degradation caused by connectivity noise, particularly in the intermediate coupling regime where noise levels are substantial. While prior works suggest that high-fidelity operation requires strongly suppressed noise regimes, maintaining such conditions under practical experimental constraints remains challenging. The authors investigate quantum gate operations in fully connected rings, examining both SWAP and general circuits. Their central finding is that fidelity can be significantly enhanced by tuning gate operation durations, with local maxima emerging even under strong noise conditions. These fidelity enhancements occur consistently across different qubit numbers and operation types. Remarkably, for specific initial states — particularly those with favorable symmetry or entanglement properties — the achieved fidelities approach quantum error correction thresholds, a critical milestone for scalable fault-tolerant quantum computation. Crucially for the AI4S theme, a supervised machine learning model that accurately predicts the optimal operation durations for new devices was developed, enabling efficient optimization without extensive experimental simulations. This represents a powerful example of AI serving as an indispensable tool for quantum hardware characterization and control.
Quantum secure communication and AI. In a complementary development, Chen and co-workers [
14] introduced the concept of controller-independent controlled bidirectional quantum secure direct communication (CICBQSDC) — a protocol that enables two legitimate users to confidentially exchange secret messages with the permission of a controller, while restricting the controller from obtaining useful information on the secret messages. The authors construct a novel protocol that avoids four critical security problems: information leakage, violation of simultaneity, uncontrolled communication, and controller eavesdropping. Unconditional security is demonstrated in the asymptotic scenario. While this work is primarily in quantum cryptography, its relevance to AI4S lies in the growing need for secure data sharing in collaborative AI-driven scientific research. As AI models are increasingly trained on distributed, sensitive datasets — including proprietary experimental data, medical information, or classified materials data — quantum-secure communication protocols may become essential infrastructure for the AI4S ecosystem.
2.6 Reinforcement learning and complex systems for autonomous experimentation
Beyond property prediction and system optimization, AI4S also leverages reinforcement learning (RL) to reveal emergent physical laws of complex collective systems and drive autonomous experimental automation. In 2024, Zhang
et al. [
15] explored the dynamics of multi-agent systems where each agent employs reinforcement learning. Their study of a reinforcement learning minority game system reveals that as AI agents learn to optimize their actions, the collective system approaches an optimal state through self-organizing oscillations that suppress herding behavior [
10]. Notably, they observed a first-order phase transition in this multi-agent system — a finding that bridges AI research with statistical physics [
15]. This work demonstrates how AI can serve not only as a tool for scientific discovery but also as a subject of scientific inquiry, revealing fundamental principles of collective behavior.
Furthermore, RL and Bayesian optimization have been widely deployed in autonomous laboratory construction. Fakhruldeen
et al. [
16] automated solubility screening and crystallization by combining commercial robots with standard laboratories utilizing a robot operating system (ROS) architecture. To conduct high-throughput experiments using robots, Burger
et al. [
17] at the University of Liverpool employed a batch Bayesian search algorithm to automate 688 experiments over the span of 8 days aimed at identifying highly active photocatalysts. These studies validate the feasibility of AI-driven closed-loop experimental paradigms for future self-driving laboratories.
3 Methodological innovations and challenges
3.1 Key Methodological approaches
Synthesizing existing studies, the mainstream machine learning methodologies applied across physical and material AI4S research can be categorized into multiple paradigmatically distinct frameworks, each with targeted application scenarios and unique technical advantages. A comprehensive summary of core methods, applicable domains, and technical merits is presented in Table 1.
3.2 Challenges
Despite significant progress, several cross-cutting challenges emerge from this body of work:
Data scarcity. Multiple papers highlight the challenge of small datasets, particularly in materials science where experimental data collection is costly and time-consuming [
18,
19]. The HNN approach and Random Forests prescreening represent two different strategies to address this limitation, but the fundamental problem remains: the data available for training AI models in scientific domains is often orders of magnitude smaller than that used in commercial AI applications.
The disconnect between physics and algorithms. Hao
et al. [
1] explicitly note the gap between physical principles and machine learning algorithms as a major limitation in LIBS analysis. This reflects a broader tension in AI4S: how to ensure that AI models are not merely correlational but incorporate physical constraints and interpretability. The field of machine-learned potentials (MLP), as comprehensively reviewed by Friederich
et al. [
19], offers valuable insights into addressing this challenge (Fig. 2). Unlike black-box regressions, successful MLP are deliberately built upon physics-inspired representations — such as symmetry functions, smooth overlap of atomic positions (SOAP), and gradient-domain learning — that encode translational, rotational, and permutational invariance, and explicitly conserve energy by predicting forces as derivatives of the potential energy surface [
5,
19]. Moreover, active learning and uncertainty quantification are seamlessly integrated to ensure that models remain reliable in unexplored regions, effectively bridging the gap between data-driven predictions and first-principles physics [
3,
5,
19]. The review further emphasizes that even the most accurate ML potentials must ultimately be coupled with deep chemical and physical insight to guide synthesis and experimental validation — a sentiment that resonates strongly with Smit and Garcia’s critique of the “data-only illusion” [
20]. Indeed, as Friederich
et al. argued, MLP are not meant to replace ab initio methods but to complement them, and their true power lies in a symbiotic integration of statistical learning and fundamental physical laws [
19]. Additionally, the gap between fundamental MLP algorithm innovation and practical large-scale materials simulation application impedes the translation of theoretical advances into real-world technological benefits.
Generalization and robustness. Several papers express concern about the generalization ability of AI models [
1,
3]. Models that perform well on training data may fail when applied to new materials, experimental conditions, or system configurations.
Massive data computation/processing and scalability. The scale of data generated by modern scientific instruments poses challenges for both storage and processing [
1,
19]. This is particularly acute in fields like spectroscopy and high-throughput materials screening via computational simulation and/or high-throughput experimentation. The training and inference of large-scale AI models consume substantial computational power, which poses a significant barrier to widespread adoption. A pressing question is how to achieve efficient AI4S applications under resource-constrained conditions — especially in edge-computing environments and real-time analytical scenarios, where both latency and energy budgets are severely limited. Addressing these systemic challenges will consolidate MLPs as a universal, reliable simulation tool for next-generation matter modelling, laying a solid foundation for fully computational, intelligent, and autonomous materials design in condensed-matter physics and materials science. Meanwhile, they also require not only algorithmic innovations (such as model compression, pruning, and distributed learning) but also strategic investments in scalable and energy-efficient hardware infrastructures tailored for scientific workloads.
4 Outlook and future directions
Current AI4S research is dominated by correlation-based pattern recognition and property prediction, while future development will focus on causal reasoning and physical mechanism interpretation. The emerging explainable artificial intelligence (XAI) and physics-informed neural networks (PINNs) will become core methodological tools, embedding physical laws, thermodynamic constraints, and quantum principles into AI model training. This paradigm shift will break the “black-box” limitation of traditional data-driven models, enabling AI systems to not only predict scientific phenomena but also explain inherent physical mechanisms, realizing the essential goal of AI-assisted scientific discovery.
Collectively, the aforementioned technical limitations and systemic bottlenecks constrain the comprehensive popularization and industrialization of AI4S technologies, indicating that pure data-driven strategies may no longer suffice for the demands of high-reliability and sustainable scientific discovery. To address these inherent defects and further unlock the innovative potential of interdisciplinary intelligent integration, the future evolution of AI4S requires paradigm upgrades from methodological optimization, hardware-algorithm co-iteration, and institutional ecosystem construction. Combined with cutting-edge research trends and unmet scientific demands, this section systematically elaborates the core developmental directions and long-term prospects of the AI4S field.
The memristor literature points to a future of tight co-design between AI algorithms and the hardware that runs them. As transformer-based models continue to grow in scale and complexity, the demands on in-memory computing hardware will intensify [
10]. This creates opportunities for novel device architectures — including 2D ferroelectric materials, halide perovskites, and 3D vertical structures — that are specifically optimized for AI workloads [
5,
6,
8]. The intersection of quantum computing and machine learning is still in its infancy, but early research advances including quantum graph neural networks and hybrid quantum-classical algorithms have demonstrated tremendous application potential. As quantum hardware matures, quantum machine learning may offer advantages for problems that are classically intractable.
Looking beyond the surveyed paper, in the long term, the ultimate vision of AI4S and MLP technology lies in fully autonomous computational and experimental materials discovery ecosystems. Integrating robust, mechanism-consistent MLP simulators with reinforcement learning optimization, high-throughput materials screening, and automated experimental synthesis workflows will establish closed-loop intelligent discovery pipelines. These autonomous systems can iteratively design material structures, perform large-scale atomistic simulations via MLPs, analyze dynamic behavioral mechanisms, refine physical hypotheses, and validate experimental results, drastically accelerating the pace of novel functional material development. Notably, the advancement of high-performance, generalizable MLPs and broader AI4S technologies relies inherently on cross-disciplinary integration spanning condensed-matter physics, materials science, quantum chemistry, and computer science. Cutting-edge interdisciplinary research — including topological insulator device physics and in-memory computing architecture integration, as well as superconducting quantum gate optimization via machine learning — exemplifies how cross-field collaboration resolves isolated technical bottlenecks. Additionally, the gap between fundamental MLP algorithm innovation and practical large-scale materials simulation application impedes the translation of theoretical advances into real-world technological benefits.
In general, the challenges of AI4S — from device physics to algorithm design to scientific interpretation — require expertise that spans physics, materials science, computer science, and beyond. The cryogenic magnetic topological insulator (MTI) work by Liu
et al. is a particularly striking example of interdisciplinary collaboration, bringing together expertise in topological materials synthesis, low-temperature transport measurements, device fabrication, and in-memory computing architectures [
21]. Similarly, the quantum gate optimization work by Fu and co-workers combines superconducting circuit physics, machine learning, and quantum information theory. Such collaborations are not merely beneficial but essential for advancing the field [
13]. It should be highlighted that Smit and Garcia’s call for coupling AI with “deep chemical insight” suggests that such autonomous systems must be grounded in domain expertise — they cannot simply be “black boxes” that optimize purely on the basis of data [
20].
Beyond the technical and conceptual hurdles discussed above, the long-term sustainability of AI4S innovation faces additional systemic challenges that demand coordinated action from policymakers, academic institutions, industry, and the broader research community. These challenges, if left unaddressed, could impede the translation of AI-driven predictions into real-world scientific and technological benefits.
Standard database deficiency. Establishing standardized, high-quality databases is a cornerstone of data-driven scientific innovation, yet it remains a major obstacle. The quality of existing databases is often compromised by inconsistencies in data sources, formats, and quality assurance protocols. Factors such as timeliness, completeness, accuracy, and reliability vary widely across datasets, making it difficult to train robust and generalizable AI models. Moreover, the lack of community-accepted standards for data collection, curation, and sharing hinders reproducibility and cross-study comparisons. For AI4S to mature, the scientific community must invest in building and maintaining curated, standardized databases that are openly accessible and interoperable across disciplines. This requires not only technical infrastructure but also sustained funding and institutional commitment.
Technical barriers and talent shortage. The rapid evolution of AI and data science tools creates a persistent skills gap. Personnel in universities, research institutes, and industrial R&D departments must possess a blend of domain-specific scientific knowledge and advanced data-processing capabilities. However, the pace of technological change often outstrips the capacity of educational and training programmes to keep up. Furthermore, the AI4S ecosystem demands interdisciplinary experts who are equally fluent in data science, machine learning, and the underlying physics or chemistry of the problem at hand. Such talent is in critically short supply globally. To address this, educational curricula must be reformed to integrate data science and AI training into traditional physical sciences programmes, while universities and research centers should foster cross-disciplinary training initiatives and industry-academia exchange programmes to cultivate the next generation of AI4S practitioners.
Legal and regulatory issues. Data-driven scientific innovation inevitably raises complex issues related to privacy, security, and intellectual property. The legal and regulatory landscapes vary significantly across countries and regions, creating uncertainty for collaborative and cross-border research. For example, the handling of proprietary or sensitive experimental data, the sharing of large datasets across jurisdictions, and the ownership of AI-generated knowledge all require clear legal frameworks. Innovation entities must navigate these diverse regulations to avoid legal risks and economic losses. In particular, cross-border data transfer — which is often essential for large-scale collaborative AI4S projects — demands harmonization of standards and secure transmission protocols. Policymakers should work towards internationally agreed guidelines that balance the free flow of scientific data with legitimate concerns over intellectual property and national security, while research institutions should proactively implement best practices for data governance and compliance.
5 Conclusion
This survey report has presented a picture of a field in rapid transformation. AI is no longer merely an auxiliary tool but is becoming integral to the scientific process across multiple domains. In spectroscopy, machine learning enables quantitative analysis that was previously unattainable. In materials science, AI predicts properties of novel compounds with remarkable accuracy. In device physics, memristors provide the hardware foundation for energy-efficient neuromorphic computing. In quantum information, graph neural networks offer new ways to understand quantum circuits.
Yet significant challenges remain. Data scarcity, the gap between physics and algorithms, generalization limitations, and the need for explainability all require continued research attention. We would call for interdisciplinary collaboration to propel the future of AI innovation. This theme resonates throughout the literature: the challenges of AI4S — from device physics to algorithm design to scientific interpretation — require expertise that spans physics, materials science, computer science, and beyond. Therefore, the most exciting opportunities lie at the intersections — between hardware and algorithms, between classical and quantum computing, and between traditional physics and data science. From the early conceptual analogy of Hopfield networks and Ising models to modern variational neural annealing and physics-informed neural networks, this interdisciplinary fusion has consistently delivered transformative methodological advances. Nevertheless, the field is still in a stage of rapid and cautious development, constrained by both technical and systemic challenges. Furthermore, as Smit and Garcia have provocatively framed it very recently, AI may have transformed image and language generation, but in materials science, “data scarcity and synthesis complexity demand a different approach. Only by coupling artificial intelligence with deep chemical insight can we turn virtual predictions into real materials” [
20]. This “data-only illusion” serves as a crucial corrective to the often-exuberant claims surrounding AI4S, reminding us that the ultimate measure of success is not predictive accuracy alone, but the ability to translate predictions into tangible scientific and technological advances.
As the AI4S field continues to evolve rapidly, the synergistic integration of physics and machine learning will remain its core driving force. The future development of AI4S lies in the symbiotic integration of artificial intelligence and fundamental scientific principles, rather than the simple replacement of domain knowledge by data-driven algorithms. The disciplinary boundary between “AI for Science” and “Science for AI” will gradually blur and merge: AI systems will accelerate the discovery of unknown physical and chemical mechanisms, while scientific domain insights will guide the design of more interpretable, robust, and discipline-adaptive AI architectures. Future breakthroughs will focus on causal mechanism mining, hardware-AI co-design, quantum machine learning innovation, and autonomous laboratory construction, accompanied by continuous optimization of institutional systems and talent training mechanisms.
In summary, AI4S represents one of the most transformative research paradigms in modern physical and material sciences. The ultimate value of AI4S innovation is reflected in tangible scientific and technological outputs, including novel functional materials, high-performance electronic devices, and advanced intelligent scientific technologies. To realise this vision, the community must embrace both the power of AI and the imperative of deep chemical, physical, and institutional scientific insights.
PDF
(1775KB)
