Composition-Weighted Symbolic Regression for General-Purpose Property Prediction

This paper introduces a composition-weighted symbolic regression framework that combines hybrid search algorithms with max/min operators to generate interpretable, analytical expressions for predicting diverse materials properties directly from chemical composition, achieving competitive accuracy against black-box models while revealing chemically meaningful elemental trends.

The Gist

Imagine you are a chef trying to figure out the exact recipe for a perfect cake. Usually, scientists trying to predict how a material will behave (like whether it conducts electricity or how hard it is) use two main approaches:

1. The "Blueprint" Approach: They look at the detailed 3D structure of the atoms (the blueprint). This is very accurate but requires knowing the blueprint, which is often missing or too expensive to build.
2. The "Black Box" Approach: They look only at the list of ingredients (the chemical formula) and feed it into a giant, complex computer brain (a neural network). This brain gives a correct answer, but no one knows how it got there. It's like the chef saying, "It tastes good," but refusing to tell you the recipe.

This paper introduces a new method called Composition-Weighted Symbolic Regression. Think of this as a smart, transparent recipe finder that only looks at the list of ingredients but still manages to write down the actual mathematical recipe for the material's properties.

Here is how it works, broken down into simple concepts:

1. The "Weighted Ingredient" Idea

Instead of just listing ingredients, the method assigns a "score" or "weight" to each element (like Carbon, Iron, or Oxygen).

• The Analogy: Imagine you are making a soup. The recipe isn't just "add carrots." It's "add 2 parts carrots, 0.5 parts salt, and -1 part sugar (because you don't want it sweet)."
• The computer learns these specific weights for every element automatically. It figures out that for a "hard" material, Iron might get a high positive score, while for a "soft" material, it might get a negative score.

2. The "Mathematical Recipe" (Symbolic Regression)

Once the computer has the ingredient weights, it doesn't just guess the answer. It searches for the actual mathematical formula that connects those weights to the final result.

• The Analogy: Instead of a black box that says "Result: 5," it writes out: Result = (Weight of Iron × 2) + (Weight of Carbon ÷ 3).
• This is called "Symbolic Regression." It finds the equation itself, making the prediction interpretable. You can read the formula and understand the logic.

3. The "Safety Guards" (Max/Min Operators)

Materials have physical rules. For example, a "band gap" (a measure of how well a material blocks electricity) can never be negative. A probability (like "chance this is a metal") must be between 0 and 1.

• The Analogy: Imagine a thermostat that has a hard stop so it can't go below freezing, or a speedometer that can't show negative speed.
• This method builds those "safety guards" directly into the math using Max and Min functions. If the math tries to calculate a negative band gap, the "Max" function acts like a floor, saying, "No, the lowest this can be is zero." This ensures the results always make physical sense.

4. The "Search Team" (Hybrid Algorithm)

Finding the perfect recipe and the perfect weights is like finding a needle in a haystack. The authors used a clever team of two searchers:

• The Explorer (Monte Carlo Tree Search): This part explores different paths, like a hiker trying different trails in a forest to find the best view.
• The Refiner (Genetic Programming): This part acts like a breeding program. It takes the best "recipes" found so far, mixes them together, and tweaks them to make them even better.
• The Coach (Gradient-Based Optimization): Once a promising recipe is found, a coach steps in to fine-tune the numbers (the weights) precisely, ensuring the math is as accurate as possible.

What Did They Find?

The authors tested this method on a standard set of material data (MatBench).

• Accuracy: It performed almost as well as the giant "Black Box" computer brains, even though it uses far fewer "parameters" (it's much simpler).
• Smoothness: When predicting properties for new mixtures of materials (like mixing two semiconductors), the "Black Box" models sometimes jump around wildly or give jagged, unrealistic results. This new method produces a smooth, continuous curve, like a well-drawn line on a graph, which is much more realistic for how materials actually behave.
• Chemical Sense: When they looked at the "weights" the computer learned, they matched real chemistry. For example, elements that are chemically similar (like those in the same column of the Periodic Table) got similar scores. The computer "rediscovered" chemical patterns on its own without being told what they were.

The Catch (Limitations)

The authors are honest about the downsides:

• Complexity: Sometimes the "recipe" the computer finds is still very complicated and hard for a human to read, even if it is mathematically explicit.
• Not Perfect: The search method is very good but doesn't guarantee it found the absolute best possible answer every time.
• Data Hungry: If you don't have enough data, the computer might get too creative and invent a complex recipe that fits the data but doesn't reflect reality (overfitting).

Summary

In short, this paper presents a tool that acts like a detective chemist. It looks at a list of ingredients, figures out the hidden mathematical rules that govern the material's behavior, and writes down a clear, logical formula. It bridges the gap between the high accuracy of complex AI and the clear understanding of traditional science.

Technical Summary

Technical Summary: Composition-Weighted Symbolic Regression for General-Purpose Property Prediction

Problem Statement
Current machine learning approaches for materials property prediction are generally categorized into structure-based and composition-based methods. While structure-based models (e.g., Equiformer, TACE) achieve high accuracy by leveraging atomic configurations, they are limited by the frequent unavailability, uncertainty, or high computational cost of structural data. Composition-based methods offer a solution by predicting properties directly from chemical formulas, enabling rapid screening. However, most existing composition-based models rely on neural networks or black-box architectures that lack physical interpretability. The central challenge addressed by this work is how to maintain competitive predictive accuracy while recovering transparent, chemically meaningful analytical relationships without relying on predefined descriptors or prior physical assumptions.

Methodology
The authors propose a composition-weighted symbolic regression framework that jointly learns analytical functional forms and task-dependent elemental weightings. The core formulation expresses a material property $P$ as:
$$P = F(x; \theta), \quad x_k = \sum_i w_{k,i} c_i$$
where $c_i$ represents the elemental composition fraction, $w_{k,i}$ are learnable elemental weights, and $F$ is an analytical function identified via symbolic regression. The variables $x$ represent composition-weighted averages of latent elemental properties.

Key methodological components include:

1. Expanded Operator Set: The search space includes standard continuous operators (exp, log, multiplication, addition) alongside non-smooth operators, specifically max and min. This inclusion allows the model to naturally enforce physical constraints, such as non-negative band gaps or bounded classification probabilities $[0, 1]$, unifying regression and classification tasks within a single symbolic formalism.
2. Hybrid Optimization Algorithm: To navigate the enlarged search space (which includes both symbolic structures and high-dimensional elemental weights), the authors employ a hybrid Monte Carlo Tree Search (MCTS) and Genetic Programming (GP) framework.
   • MCTS-GP Integration: The method combines the directed exploration of MCTS with the "stage-jumping" capabilities of GP. Unlike previous implementations that store candidate queues at many nodes, this approach retains the global expression queue only at the root node, performing all genetic operations (mutation, crossover) on this shared population to reduce memory overhead.
   • Gradient-Based Refinement: For continuous parameter optimization (elemental weights $w$ and symbolic coefficients $\theta$), the framework utilizes the Limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) algorithm. A multi-start strategy is employed to handle non-smoothness introduced by max/min operators, ensuring robustness against local minima.
   • Parallelism: Both the GP and MCTS stages are parallelized to improve computational efficiency, with batch processing for expression generation and parameter optimization.

Key Results
The framework was evaluated on three representative MatBench tasks: band gap prediction (regression), metallicity classification, and glass formation classification.

• Benchmark Performance: The model achieved competitive accuracy relative to state-of-the-art black-box models (including CrabNet, MODNet, and large language models like Darwin and GPTChem) while utilizing significantly fewer trainable parameters (approx. $10^2$ vs. $10^6$ to $10^9$ for neural networks).
  • Band Gap: Mean Absolute Error (MAE) of 0.471, compared to 0.287 for the 7B-parameter Darwin model and 0.331 for CrabNet.
  • Metallicity: ROC-AUC of 0.873, comparable to MODNet (0.916) and CrabNet (not reported).
  • Glass Formation: ROC-AUC of 0.816, comparable to MODNet (0.960) and RF-SCM (0.859).
• Interpretability and Periodic Trends: The model successfully recovered explicit analytical expressions (e.g., $F_{gap} = x_1 \exp[-\exp(\max(x_2, \min(x_0, x_1)))]$). The learned elemental weights exhibited chemically meaningful periodic trends. For instance, halogens displayed a specific weight pattern consistent with their role in stabilizing insulating environments, while transition metals showed patterns associated with metallic bonding.
• III–V Semiconductor Alloys: When applied to predict band gaps for III–V ternary alloys, the symbolic model produced smooth, continuous composition-dependent trends. In contrast, neural network-based models (Darwin, CrabNet, MODNet) exhibited discontinuities or fluctuations in regions with sparse training data. The symbolic approach provided physically consistent interpolation, correctly reproducing global trends such as the band gap decrease from AlAs to InSb.

Significance and Claims
The paper claims to provide a scalable and interpretable route for materials discovery and property screening. Its primary significance lies in:

1. Unifying Regression and Classification: By incorporating max/min operators, the framework handles bounded outputs and physical constraints (e.g., non-negativity) directly within the learned expression, eliminating the need for task-specific output layers.
2. Data-Driven Functional Discovery: The method learns both the functional form and elemental representations directly from data, avoiding the bias of hand-crafted descriptors.
3. Physical Consistency: The resulting closed-form expressions ensure smooth behavior across continuous composition spaces, offering a distinct advantage over black-box models for interpolation and extrapolation in data-sparse regimes.

Limitations
The authors acknowledge several limitations:

• Interpretability vs. Complexity: While expressions are explicit, highly accurate solutions may be algebraically complex, requiring further analysis to extract physical insights.
• Optimization Approximation: The hybrid MCTS-GP strategy does not guarantee global optimality, and the gradient-based stage is inherently local.
• Overfitting: In low-data regimes, the flexibility of symbolic regression may lead to overly complex expressions that fit noise rather than underlying physical trends.
• Functional Space: The current operator set may be insufficient for strongly multiscale or sharply discontinuous phenomena, such as complex phase-boundary behaviors.