A Decade of Data-Driven Materials Engineering: Progress, Remaining Gaps, and Unlearned Lessons

Machine learning has transformed materials engineering through graph neural networks, equivariant architectures, generative models, and early autonomous laboratories. These advances enabled accurate property prediction, data-efficient force fields, and initial inverse design of inorganic crystals, supported by community databases such as the Materials Project and JARVIS. Generative frameworks now propose novel structures, while closed-loop platforms demonstrate early integration of prediction, synthesis, and robotic experimentation. However, major gaps remain: weak extrapolation beyond training distributions, inadequate handling of long-range interactions, poorly calibrated uncertainty quantification, limited synthesis prediction, underrepresentation of disordered materials, and insufficient experimental validation. Persistent unlearned lessons include benchmark biases that inflate generalization performance, poor reproducibility practices, misuse of invariant models for tensor properties, omission of random baselines in active learning, weak validity filtering in generative workflows, and near-absent failure reporting. This review critically assesses a decade of progress and shortcomings, grounded in 38 peer-reviewed publications. It highlights that hype around foundation models, generative inverse design, and autonomous labs often exceeds demonstrated impact. Realizing the field’s potential requires prioritizing extrapolation-focused architectures, mandatory experimental validation, shared failure registries, and stronger evidentiary standards to accelerate reliable discovery of functional materials.

Recent years have seen remarkable advancements in data-driven materials engineering.From early GNNs like CGCNN and SchNet [1-3] to equivariant architectures [4, 5] to generative models [6, 7], the field has transformed [8]. This review assesses progress over the decade, identifies remaining gaps, and highlights unlearned lessons — insights that have been repeatedly demonstrated but not yet absorbed into common practice. We cover architectures, property prediction, force fields, generative models, active learning, and experimental validation.

Data-driven approaches have complemented traditional density-functional theory by enabling rapid screening of vast compositional spaces that would be intractable with first-principles methods alone [9, 10]. The introduction of graph-based representations allowed models to learn directly from atomic connectivity, eliminating the need for many hand-crafted descriptors that previously limited transferability [1, 11, 12]. Community-scale databases quickly became essential infrastructure, providing standardized training and test sets that facilitated direct comparison of methods [13, 14]. By the end of the decade, thousands of publications cited these resources as the foundation for model training and benchmarking.

Progress was not linear. The period 2017–2019 established the dominance of GNNs for crystalline systems and laid the groundwork for ML force fields [1, 9]. The subsequent years 2020–2022 brought the recognition that physical symmetries must be explicitly encoded, leading to equivariant architectures that achieved state-of-the-art accuracy with far smaller datasets [4]. From 2023 onward, the focus shifted toward generative modeling and closed-loop discovery, with diffusion models proposing entirely new crystal structures and early autonomous platforms integrating prediction with synthesis [6, 7, 15, 16]. These developments were accompanied by growing interest in multimodal models that incorporate text from the literature alongside structural data [10, 17, 18].

Despite the impressive volume of research, the field remains fragmented. Many studies report incremental improvements on narrow tasks while overlooking broader challenges such as extrapolation and synthesis [19, 20]. Reviews published throughout the decade repeatedly flag the same issues—benchmark biases, reproducibility shortfalls, and insufficient experimental grounding—yet these concerns have not translated into widespread changes in practice [21-24]. This review therefore adopts a dual mandate: to synthesize what has been achieved and to confront what has not.

Figure 1 maps the review’s central argument by showing how decade-scale methodological advances coexist with unresolved gaps, unlearned lessons, hype inflation, and the infrastructural reforms required for reliable discovery.

2017–2019: The GNN revolution The period 2017–2019 witnessed the rapid establishment of graph neural networks as the dominant architecture for materials property prediction. CGCNN and SchNet demonstrated that crystal graphs could directly encode atomic environments and predict formation energies, band gaps, and elastic properties with accuracy rivaling DFT [1, 2]. MEGNet further showed that the same framework could be extended to multiple fidelities and different target properties within a single model [2]. These early successes were enabled by the Materials Project, which provided a standardized, publicly accessible benchmark dataset of over 100 000 inorganic compounds [13, 14]. Within a few years, the Materials Project API became the de-facto standard for training and testing, allowing direct head-to-head comparisons across research groups worldwide.

Simultaneously, ML force fields began to mature. Approaches such as DeepMD and GAP illustrated how neural networks could reproduce ab initio potential-energy surfaces at a fraction of the computational cost, opening the door to large-scale molecular-dynamics simulations of realistic materials systems [9]. By the end of this phase, GNNs had proven their ability to capture local atomic interactions and were routinely applied to high-throughput screening of novel compositions [10, 25, 26].

2020–2022: Equivariance and data efficiency The next wave of progress centered on the explicit incorporation of physical symmetries. Equivariant GNNs such as those underlying NequIP, MACE, and Allegro achieved state-of-the-art accuracy on force-field benchmarks while requiring significantly less training data than their invariant predecessors [4, 5]. The recognition that tensor properties (piezoelectric coefficients, dielectric tensors) require equivariant representations became widespread [5]. These models demonstrated that respecting rotational, translational, and permutational symmetries is not merely an aesthetic choice but a practical necessity for reliable extrapolation within known chemical spaces [4].

Active learning also gained traction during this period [27]. Lookman et al. and Cao et al. showed that uncertainty-guided acquisition could accelerate the discovery of high-performance solders and other functional materials [28, 29]. Multi-fidelity strategies further improved data efficiency by intelligently combining cheap low-accuracy calculations with sparse high-accuracy results [6, 28]. By 2022, equivariant architectures were routinely cited as the new baseline for any force-field or property-prediction task involving periodic systems.

2023–2026: Generative models and foundation models The final phase of the decade was defined by the rise of generative methods and early foundation-model efforts. Diffusion models proved capable of generating stable crystal structures conditioned on composition or desired properties, offering a route to true inverse design [6, 15]. Specialized generative frameworks for inorganic materials reached publication in high-impact venues, demonstrating the feasibility of proposing candidates outside known databases [6, 7]. Large language models began to be integrated for literature mining and hypothesis generation, although their quantitative predictive power remained limited [10, 18].

Autonomous laboratories emerged as proof-of-concept platforms that combined prediction, synthesis planning, and robotic experimentation in closed loops [7]. Multi-fidelity and multimodal learning further expanded the scope of what could be learned from heterogeneous data sources [6, 28]. By 2025–2026, the community had produced the first demonstrations of end-to-end discovery pipelines that moved from idea to validated material within a single workflow.

Key milestones timeline

Collectively, these advances have reduced the time from concept to candidate material from years to months in selected domains and have generated thousands of new hypothetical compounds for further study [1, 32]. The decade has therefore delivered tangible methodological progress that no longer requires justification; the challenge now lies in translating these tools into routinely validated, application-ready materials.

Current models excel as interpolators yet falter as extrapolators, with prediction errors surging under new elements, unseen prototypes, or extreme out-of-distribution conditions [19, 20, 33]. No architecture has achieved fundamental advances in generalization for materials, causing data-efficiency gains within familiar chemical spaces to collapse for novel compositions [18, 23]. This limitation compounds with the reliance of most GNNs on short-range cutoffs that overlook long-range electrostatic, dispersion, and polarization effects critical to ionic conductivity, polar crystals, and collective phenomena in electrolytes [5, 34], an issue persisting without standard solutions via extended message passing or hybrid potentials [4]. Such shortcomings extend to miscalibrated uncertainty estimates that rarely separate epistemic from aleatoric components, where even conformal methods sacrifice sharpness and thus practical utility for screening [18, 22, 35]. Synthesis prediction remains equally intractable, as generative approaches yield metastable or inaccessible structures while language models offer unverified recipes [6, 7, 10, 31, 36]. Progress further concentrates on ordered crystals, leaving disordered systems—amorphous solids, glasses, and defects—underrepresented amid acute size-transferability challenges and absent community benchmarks [23, 37]. Finally, validation against DFT data rather than experiments propagates functional errors, with few studies completing the discovery-to-measurement cycle and thereby constraining real-world impact [7, 24, 30].

Benchmark biases in standard datasets, such as the Materials Project’s emphasis on stable, ordered, and cubic structures, cause random train–test splits to inflate generalization estimates by leaking chemically similar compounds across partitions [21, 23]. This structural artifact undermines reliability, even as calls for multiple test splits and cross-dataset validation have gone largely unheeded after a decade [18]. Compounding the issue, reproducibility suffers from widespread omission of code, data, and environments, while model cards detailing intended use and limitations remain rare despite repeated recommendations [18, 22, 23]. Architecture choices introduce further distortions: invariant GNNs inherently fail on tensorial properties such as piezoelectricity by forcing equivariant quantities toward zero, yet invariant models continue to be misapplied where symmetry demands equivariance [4, 5]. In active learning, complex acquisition functions often yield only marginal gains over random sampling once proper baselines are included, yet many studies omit these baselines and thereby overstate acceleration [28, 29, 35]. Generative models similarly require rigorous post hoc filtering, as most proposed structures prove invalid, unstable, or inaccessible, with papers still highlighting “novel materials” while reporting validity rates as low as 10 % [6, 7, 36]. Finally, the persistent reliance on DFT validation propagates systematic functional errors, and although experimental confirmation constitutes the true benchmark, the community has yet to move beyond DFT-level agreement in the majority of cases [24, 30].

Although foundation models have been heralded for zero-shot generalization across materials properties and seamless literature-to-prediction pipelines, their errors remain an order of magnitude higher than those of fine-tuned GNNs on equivalent tasks, confining them largely to summarization and hypothesis generation rather than quantitative prediction [10, 18]. Generative models, similarly promoted as enablers of routine inverse design for synthesizable compounds, continue to produce mostly invalid or unstable candidates with validity rates frequently below 20 % even after filtering, thus serving better for chemical space exploration than for delivering experimentally confirmed materials [6, 7, 36]. This gap in delivery extends to autonomous laboratories, which, despite visions of closed-loop discovery replacing human scientists on accelerated timescales, remain costly, restricted to simple syntheses, and heavily dependent on human intervention for non-routine processes, limiting their current industrial scalability [7, 18]. Equivariant GNNs, positioned as near-universal solutions for force fields and tensor properties, still struggle with extrapolation and long-range interactions without supplementary modules, rendering them the strongest available yet insufficient for the most challenging out-of-distribution regimes [4, 5, 34]. Finally, active learning’s promised dramatic reduction in experimental burden materializes as only modest improvements over rigorously controlled random sampling in many domains, underscoring that its advantages, while genuine, are incremental rather than transformative [28, 29, 35].

Table 1 shows that many of the field’s most persistent weaknesses arise from a recurring mismatch between high-visibility claims and the evidentiary standards actually required for reliable, experimentally grounded discovery.

Data-driven materials engineering has evolved in parallel with machine learning applications in neighbouring disciplines, yet it retains distinct challenges and strengths. Comparison with drug discovery is particularly instructive. Molecular machine learning benefited from early, large-scale benchmarks such as QM9, ZINC, and MoleculeNet that provided standardized, diverse datasets and clear evaluation protocols [23, 24]. These resources enabled rapid iteration and community-wide progress on tasks such as property prediction and molecular generation. Materials science, by contrast, must contend with periodicity, thermodynamic stability under realistic conditions, and the additional layer of synthesizability—factors absent or far less constraining in small-molecule design [20, 23]. Nevertheless, materials researchers can adopt the rigorous cross-validation and multi-metric reporting practices that have become routine in drug discovery, especially for out-of-distribution generalization [24].

The comparison with computer vision highlights another set of differences. Computer vision enjoys enormous, curated datasets (ImageNet and its successors) and standardized benchmarks that have driven decades of architectural innovation [18]. Materials datasets remain smaller, heavily biased toward stable crystalline compounds, and lack the same level of curation or diversity [21, 23]. Transfer learning strategies successful in vision—pre-training on massive image corpora—do not translate directly because atomic-scale structures do not share the same statistical regularities as natural images [18]. Domain-specific adaptations remain necessary.

Despite these limitations, materials artificial intelligence offers advantages not yet fully realized in other fields. The explicit incorporation of physics—through equivariant architectures that respect rotational, translational, and permutational symmetries—has produced models that conserve energy and momentum by construction [4, 5]. This physics-informed approach contrasts with purely data-driven methods in drug discovery or computer vision, where physical consistency must often be imposed post hoc [4]. Integration of domain knowledge, such as crystal symmetry or known phase diagrams, further improves data efficiency and interpretability in ways that are less common in molecular or image-based machine learning [5, 34]. Multi-fidelity strategies that combine cheap approximate calculations with sparse high-accuracy data have also matured faster in materials than in many other domains [6, 28].

Overall, materials informatics has learned from the benchmarking culture of drug discovery and the architectural innovations of computer vision, yet it has developed unique strengths in embedding physical priors and handling structured, periodic data. The next decade will determine whether these strengths can be leveraged to close the remaining performance gaps or whether the field will continue to lag behind its more mature neighbours in terms of standardization and reproducibility [18, 22].

The coming decade demands a shift from incremental gains on familiar tasks toward resolving the structural limitations outlined here. Future architectures must prioritize extrapolation by moving beyond local cutoffs and interpolation in known chemical spaces, integrating explicit long-range modules and physics-informed priors capable of handling unseen elements, novel prototypes, and extreme conditions [5, 19], with causal modelling frameworks offering a compelling path to disentangle correlation from causation [18]. This progress hinges on deeper experimental integration: although autonomous platforms have shown initial closed-loop promise, they remain narrow in scope, requiring the community to elevate laboratory validation as the default and insist on experimental confirmation for any significant discovery claim rather than DFT agreement alone [7, 24, 30]. Hybrid human–robot systems linking predictive models to high-throughput synthesis will prove indispensable in this transition [7]. Such grounding naturally extends to uncertainty-aware decision-making, where quantification must advance from miscalibrated estimates to conditional coverage guarantees that deliver actionable confidence bounds for prioritizing synthesis [18, 22, 35]. Complementing this, shared failure databases and dedicated negative-results sections would institutionalize the documentation of unproductive pathways, invalid structures, and collapsed extrapolations, transforming repeated mistakes into collective knowledge [18, 24, 32]. Finally, truly multimodal models that fuse literature text, structural data, spectroscopy, and process parameters within one framework could harness decades of experimental insight to inform generative design and active learning, thereby closing the historical-to-forward prediction loop [10, 18]. These directions are inherently interdependent: extrapolation without experimental grounding falters, uncertainty quantification lacks impact absent failure reporting, and multimodal integration yields limited value without enforced reproducibility standards [22, 23]. Coordinated community effort, rather than isolated advances, will therefore define meaningful progress.

For individual researchers the priorities are clear: shift focus from interpolation to extrapolation, validate every major claim against experiment rather than DFT alone, and report failures alongside successes [19, 24, 32]. Model cards that document intended use, limitations, and training distributions should become mandatory supplementary material [22].

Benchmark designers must create challenging out-of-distribution test sets that span new elements, unseen prototypes, and disordered systems [21, 23]. Multiple test splits and cross-dataset validation should be required, and every benchmark release must include uncertainty-reporting protocols [18]. Random baselines must be reported for all active-learning studies [28, 35].

Journals should enforce reproducibility standards by mandating code and data deposition, model cards, and experimental validation for any discovery claim [7, 22]. Dedicated negative-results sections would accelerate collective learning and reduce redundant effort [24, 32]. Reviewers must insist on validity, novelty, and stability metrics for generative models rather than accepting headline claims based on unfiltered output [6, 36].

Funders have a pivotal role in building infrastructure. Support for community databases, standardized OOD benchmarks, and shared failure registries will yield returns far greater than additional isolated projects [18, 23]. Priority should be given to extrapolation research, autonomous-laboratory scaling, and multimodal integration that connects literature with computation [7, 10]. Long-term funding for reproducibility platforms and negative-results dissemination will embed the unlearned lessons of the past decade into standard practice [22].

Collectively, these recommendations shift the field from volume-driven publication to quality-driven discovery. If implemented, they will ensure that the next decade of data-driven materials engineering is defined not by hype cycles but by reproducible, experimentally validated advances that deliver real functional materials.

Near term (1–3 years): the community should establish standardized out-of-distribution benchmarks, mandate equivariant GNNs as the default for tensorial and force-field tasks, and require experimental validation for any new material claim [4, 21, 24]. Reproducibility checklists and model cards should become submission requirements across target journals [22].

Medium term (3–7 years): breakthroughs in extrapolation architectures should enable reliable predictions for new element combinations and disordered systems [19, 37]. Uncertainty quantification should reach the level of conditional coverage guarantees usable in risk-aware discovery campaigns [35]. Autonomous laboratories should scale to handle moderately complex syntheses, demonstrating closed-loop discovery of at least one new functional material per platform [7].

Long term (7–10 years): generalizable materials AI should integrate multimodal literature mining, physics-informed generative models, and experimental feedback into unified discovery platforms [6, 10, 18]. Experimental validation should become the norm for the majority of published predictions rather than the exception [24].

Success criteria for 2036 are concrete and measurable. Models must predict properties of novel element combinations with mean absolute errors below 0.1 eV/atom on formation energies and comparable accuracy for electronic and mechanical tensors [5, 34]. Autonomous laboratories should routinely discover and validate new materials in fewer than 100 targeted experiments [7]. Failure reporting must be standard practice, with every major study documenting unsuccessful pathways in supplementary materials or dedicated repositories [18, 32]. Benchmark biases should be routinely mitigated through multiple test splits and cross-dataset evaluation [21, 23].

Achieving these criteria will require the field to absorb the unlearned lessons of the past decade and to treat extrapolation, uncertainty, synthesis, and experimental grounding as core rather than peripheral concerns. The path forward is therefore not merely technical but cultural—moving from a literature that celebrates incremental interpolation to one that demands reproducible, generalizable, and experimentally verified discovery.

A decade of data-driven materials engineering (2017–2026) has seen remarkable progress: GNNs, equivariant architectures, generative models, and autonomous laboratories have transformed screening and hypothesis generation. Yet remaining gaps in extrapolation, long-range interactions, uncertainty quantification, synthesis prediction, disordered materials, and experimental validation continue to limit impact. Unlearned lessons—pervasive benchmark biases, inadequate reproducibility infrastructure, continued misuse of invariant architectures for tensor properties, omission of random baselines in active learning, insufficient validity filtering in generative workflows, and the near-absence of failure reporting—have been repeatedly identified but not yet absorbed into common practice. Critical assessment shows that hype surrounding foundation models, generative inverse design, and autonomous laboratories frequently exceeds demonstrated quantitative impact when judged against rigorous baselines.

The next decade must therefore prioritize extrapolation-focused research, mandatory experimental validation, shared failure registries, and realistic expectations. Only by confronting these gaps and internalizing the unlearned lessons can the field deliver on its promise to accelerate the discovery of functional materials for energy, sustainability, and advanced manufacturing. Data-driven materials engineering stands at a crossroads: the methodological foundations are in place, but the cultural and infrastructural shifts required for genuine impact remain incomplete. The literature of 2017–2026 provides both the evidence of progress and the roadmap for addressing its shortcomings. The choice to follow that roadmap now rests with the community.

None

None

None

None

Mohamed Salah, Youssef Karim, Ahmed Nabil, Mahmoud Adel & Karim Hassan contributed to this work.

Authors and affiliations

Department of Computational Materials Systems, Faculty of Engineering, Cairo University, Cairo, Egypt
Mohamed Salah, Youssef Karim & Mahmoud Adel

Department of Materials Data Analytics, Faculty of Technology, Ain Shams University, Cairo, Egypt
Ahmed Nabil & Karim Hassan

Corresponding author

Correspondence to Mohamed Salah

Open Access The author(s) retain copyright. This article is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. It may be shared and adapted for non-commercial purposes with appropriate attribution, an indication of changes, and distribution of adaptations under the same license. Third-party material may be subject to separate terms identified in its credit line. View the license at https://creativecommons.org/licenses/by-nc-sa/4.0/.