Materials science is inherently a field defined by data scarcity. Unlike image recognition or natural language processing, where tens of thousands of data points can be readily acquired, high-purity experimental data in materials science is a precious asset that demands immense costs and time. While deep learning models navigate toward solutions by tuning millions of parameters, materials scientists face the daunting challenge of identifying core variables that govern physical laws from as few as several dozen samples.
To overcome the limitations of such “Small Data” and derive clear mathematical expressions for complex physico-chemical correlations, Compressed Sensing and the SISSO (Sure Independence Screening and Sparsifying Operator) method have emerged as powerful technical solutions.
Originally developed at the Fritz Haber Institute (FHI) of the Max Planck Society, this technique focuses on the principle of sparsity—the idea that only a handful of critical physical quantities ultimately determine a material’s properties, rather than an exhaustive set of all possible data.
The SISSO framework identifies the physical mechanisms governing material properties through a two-fold process: the SIS (Sure Independence Screening) stage, which rapidly filters out highly correlated elements from a vast pool of candidate variables, and the SO (Sparsifying Operator) stage, which employs compressed sensing algorithms to determine the optimal mathematical expression using the minimum number of variables. Moving beyond “black-box” AI that merely provides predicted values, this methodology offers researchers interpretable physical insights, and its academic validity has been rigorously proven through numerous high-impact publications.
Compressed Sensing: A Mathematical Solution for Reconstructing the Whole from Minimal Information
Compressed Sensing is a concept that originally revolutionized the field of signal processing. It is based on the mathematical ingenuity that if a signal possesses sparsity, the original information can be perfectly reconstructed using far fewer data points than the minimum number of measurements prescribed by the Nyquist-Shannon sampling theorem. This suggests that focusing on the structural characteristics of information, rather than its sheer volume, enables much more efficient data reconstruction.
Applying this principle to materials science allows us to start with the powerful assumption that among tens of thousands of candidate variables (descriptors) that determine a material’s complex properties, only a few key variables actually govern the phenomenon. In this context, compressed sensing refers to the sophisticated exploration of high-dimensional parameter spaces to effectively extract these few core variables. It helps select a model that minimizes error while maintaining the simplest mathematical structure among hundreds of millions of variable combinations. By compressing intricately entangled high-dimensional data into essential low-dimensional physical quantities, it serves as a powerful mathematical compass that allows access to physical reality even in data-scarce environments.
This compressed sensing-based approach has distinct differences from Principal Component Analysis (PCA), a traditional dimension reduction technique. PCA creates new virtual axes (Principal Components) by linearly combining existing variables to preserve as much of the data’s variance as possible. However, in this process, the original physical variables become complexly scrambled, making it difficult to interpret the actual physical meaning of the resulting components.
In contrast, instead of creating and mixing new variables, compressed sensing selects the original variables themselves that best explain the material properties from a vast array of actual physical quantities. While PCA is a method of projecting dimensions while maintaining the overall shape of the data, compressed sensing is a sparse selection approach that isolates only the key variables where physical causality remains intact among trillions of candidates.
Through this, researchers can clearly identify the specific physical variables and the mathematical relationships hidden behind the results presented by artificial intelligence. This transcends simple statistical prediction, providing a robust theoretical foundation that can be directly applied to actual materials design.
The SISSO Method: Designing Useful Formulas by Recombining Materials’ Genes
The SISSO algorithm can be described as the perfect optimization of this compressed sensing philosophy for the practical field of materials discovery. Moving beyond a passive approach of simply selecting one of the given variables, SISSO implements an active mechanism that mathematically combines basic physical quantities to self-design the optimal descriptor. This process unfolds through a sophisticated three-stage pipeline.
The first stage is Feature Construction. Using basic physical quantities obtained through experiments or calculations—such as atomic number, ionization energy, and lattice constants—as raw materials, the system systematically applies mathematical operators including basic arithmetic, exponents, logarithms, and trigonometric functions. This process generates a pool of candidate complex variables numbering in the hundreds of millions. It is akin to creating a new combination of languages to most accurately describe the complex properties of a material.
Crucially, during this stage, SISSO does not merely list random mathematical combinations; it is constrained to perform only physically consistent operations. For example, it excludes the addition or subtraction of physical quantities with different units, instead focusing on creating meaningful new dimensions through multiplication or division. Specifically, it filters out “physically impermissible” operations—such as adding atomic volume to ionization energy—and ensures that only descriptors with physical plausibility are constructed. This is the core mechanism that ensures the formulas derived by the AI transcend simple curve-fitting and possess the validity required to align with actual physical laws.
The second stage is the SIS (Sure Independence Screening) process. From the hundreds of millions of generated candidate variables, it selects only the most promising candidates that show a high correlation with the target property (e.g., bandgap, thermal conductivity, formation energy, etc.). This acts as a high-performance filtering step that removes noise from the vast search space and narrows the scope to meaningful variables, thereby maximizing the efficiency of the subsequent precision analysis.
The final stage is the SO (Sparsifying Operator). A compressed sensing algorithm based on L0L_0L0 or $L_1$ regularization is applied to the candidates filtered in the previous stage. The essence of this step is to secure maximum explanatory power using the minimum number of variables. By identifying the simplest combination of variables that minimizes error, the optimal mathematical formula determining the material’s characteristics is finalized. In this way, SISSO summarizes the core physical laws hidden within complex data into a single, clear equation and delivers it to the researcher.
Why SISSO?: A White Box, Not a Black Box
While today’s rapidly advancing deep learning models hold an overwhelming advantage in predictive performance for complex non-linear data, they possess the well-known “Black Box” characteristic—making it difficult to discern the internal logic used to reach a conclusion. In a neural network structure where millions of parameters are intertwined, it is virtually impossible to physically interpret the cause of a specific result. This often acts as a critical hurdle in academic fields like materials design, where safety and reliability are the highest priorities.
In contrast, the SISSO method distinguishes itself by aiming for a concise mathematical form (y=f(x)) that humans can intuitively understand. This “White Box” model provides researchers with fundamental value that goes beyond mere predictive values. By examining the variables finally selected by the model and the mathematical relationships between them, researchers can reverse-engineer and infer the core physical and chemical mechanisms governing a material’s properties. For instance, if a specific catalyst’s activity is explained by a combination of an element’s electronegativity and its lattice constant, a researcher can establish new principles for materials design based on that insight.
Ultimately, this approach ensures that artificial intelligence does not remain a mere tool for delegated computation, but evolves into a “Digital Collaborator” that discovers and proposes unknown physical laws hidden within vast datasets. Researchers can use the formulas derived by AI to verify or refine their hypotheses. This is the path toward completing a true Materials DX (Digital Transformation)—one where data-driven insight and traditional physical theory are seamlessly integrated.
Conclusion: Quality Over Quantity, and the Insight of Algorithms
The true value of Materials DX (Digital Transformation) does not lie simply in amassing vast amounts of big data. Rather, the core competency lies in acknowledging the inherent data scarcity of materials science and possessing the ability to identify key variables that pierce through to the physical essence—generalizing them into universal formulas that anyone can understand.
From this perspective, Compressed Sensing and the SISSO method act as powerful filters that distill the simplest truths from a complex web of countless material parameters. This mathematical approach—self-designing optimal descriptors with physical plausibility among trillions of combinations and deriving them as white-box equations—elevates artificial intelligence from a mere predictor to a discoverer of new physical laws.
These technical innovations will do more than just drastically shorten the development cycle for new materials by moving away from traditional trial-and-error methods. They will serve as the most sophisticated and reliable compass for exploring unknown material territories that humanity has yet to enter due to experimental limitations. Ultimately, techniques like SISSO, which confront and overcome data limitations through mathematical insight, will undoubtedly become the essential technological assets for leading the future of materials science.