Why is data scarcity and quality the biggest roadblock?
The single most cited barrier across these studies is the lack of large, high-fidelity datasets. Machine learning models are data-hungry, but generating reliable materials data—whether through experiment or high-level quantum mechanics calculations—is slow and expensive. One paper bluntly states that the data landscape for many properties of interest is 'both scarcely populated and of dubious quality' [2]. This means models are often trained on small, noisy, or inconsistent datasets, which limits their accuracy and generalizability.
The problem is compounded by the fact that different computational methods produce conflicting results. Researchers are exploring workarounds, such as using 'consensus across functionals' in density functional theory (DFT) to improve data reliability [2], but this adds complexity. Another study highlights that while databases like the Materials Project (containing band structure data for over 63,000 materials) are a valuable resource [6], even these large repositories may not cover the specific properties or conditions needed for a given discovery task. The core issue remains: you cannot train a reliable AI without reliable data, and generating that data at scale is a fundamental scientific challenge.
Why do AI predictions so rarely lead to a real material?
A major evidence gap is the disconnect between predicting a material's stability on a computer and actually making it in a lab. AI models can screen millions of hypothetical compounds, but they often fail to account for the complex thermodynamics and kinetics of real-world synthesis. One review notes that 'examples of the data-guided discovery of entirely new, never-before-reported compounds remain limited' [7], suggesting that most AI-discovered candidates are either already known or are thermodynamically unstable under practical conditions.
This gap is partially addressed by models that predict formation energy—a key indicator of whether a compound can be made. The same study discusses how predicting the 'convex hull' (a thermodynamic stability diagram) is a critical step, and that models have succeeded in 'uncovering new DFT-stable compounds and directing materials synthesis' [7]. However, this is a far cry from routine, reliable discovery. The challenge is that even a thermodynamically stable prediction may require exotic synthesis conditions (e.g., extreme pressure [5]) that are not practical for widespread use. The AI can point the way, but the path from prediction to a physical sample remains poorly mapped.
How does the 'black box' problem hold back progress?
Even when AI models work, scientists often cannot fully trust or learn from them because the models are opaque. A key paper argues that machine learning results have been 'limited' because of its shortcomings 'in tasks requiring interpretation' [3]. This is a critical evidence gap: if a model predicts a new high-performance thermoelectric material but cannot explain *why* (e.g., which atomic features drive the high Seebeck coefficient), it provides little scientific insight and makes it hard to generalize the finding to other systems.
This is not an insurmountable problem, but it is an active area of research. Some studies show that feature analysis of trained models can 'deepen our understanding of material structure–property relationships' [1], and that weight analysis can help 'understand the sample characteristics in depth' [1]. For instance, in discovering altermagnetic materials, an AI search engine using a graph neural network 'performs much better than human experts' [4], but the paper focuses on the *what* (50 new materials) more than the *why*. The field is caught in a trade-off: the most powerful models (like deep neural networks) are often the least interpretable, while simpler, more transparent models may miss subtle patterns. Closing this interpretability gap is essential for building scientific confidence and enabling true knowledge discovery, not just pattern matching.
About These Sources
This answer is built on 8 peer-reviewed studies — published from 2021 to 2025, 3 from 2024 or later, 6 in Q1 journals, collectively cited 300 times — selected as the most relevant from 9 studies that passed quality screening, drawn from 37 papers retrieved from a database of over 500 million.
Sources used in this answer
Perspective on machine learning in energy material discovery
Highlights that while machine learning reduces the cost of high-throughput screening for energy materials (e.g., predicting band gaps), the cost of ab initio calculations remains high for complex systems, and effective data management is a key challenge.
Audacity of huge: overcoming challenges of data scarcity and data quality for machine learning in computational materials discovery
Identifies the core problem of data scarcity and quality, stating that the data landscape for many properties is 'both scarcely populated and of dubious quality', and discusses strategies like consensus across DFT functionals to mitigate this.
Materials Discovery With Machine Learning and Knowledge Discovery
Argues that machine learning results in materials science have been limited due to its shortcomings in tasks requiring interpretation, and calls for a shift toward machine-generated knowledge discovery.
AI-accelerated discovery of altermagnetic materials
Demonstrates a successful AI search engine that discovered 50 new altermagnetic materials (including 4 i-wave types), outperforming human experts, but the focus is on prediction rather than explaining the underlying physical mechanisms.
Advances in high-pressure materials discovery enabled by machine learning
Reviews machine learning-assisted crystal structure prediction under high pressure, noting that traditional approaches face challenges with computational efficiency and scalability, which ML helps to address.
Machine-Learning-Assisted Materials Discovery from Electronic Band Structure
Uses band structure data from the Materials Project (63,588 materials) to train ML clustering algorithms, showing that large databases are a valuable resource but require careful feature engineering and noise reduction.
Materials discovery through machine learning formation energy
States that examples of data-guided discovery of entirely new, never-before-reported compounds remain limited, and focuses on predicting formation energy as a critical step for determining synthetic accessibility.
Machine learning for materials discovery: Two-dimensional topological insulators
Reports a machine learning strategy that is 10x more efficient than trial-and-error for discovering 2D topological insulators, discovering 56 non-trivial materials, but notes that detailed investigation of each ML component is crucial for different results.