Causal Feature Selection for Materials Optimization: Identifying Causal Process Parameters via Machine Learning and High-Dimensional Hypothesis Testing

High-throughput experimentation (HTE) in materials science generates vast, high-dimensional datasets relating synthesis parameters to material properties. While machine learning (ML) models excel at predicting properties from these parameters, they often fail to distinguish causal drivers from merely correlated confounders. This limits rational materials design, as "feature importance" scores from standard ML models can mislead experimentalists into optimizing non-causal variables.We propose a framework to move from correlation to causation in high-dimensional HTE data. Our approach integrates Double/Debiased Machine Learning (DML) to estimate the unconfounded causal effect of each synthesis parameter, one at a time, while controlling for all other parameters as confounders. We then apply the Benjamini-Hochberg procedure to the resulting p-values, enabling high-dimensional hypothesis testing with rigorous control of the False Discovery Rate (FDR).Through a comprehensive simulation study designed to mimic the p >> n and confounded data structure of HTE, we demonstrate our framework's utility. We show that baseline ML models (e.g., Lasso, Random Forest) consistently misidentify non-causal but confounded parameters as "important." In contrast, our DML + FDR approach robustly recovers the true causal parameters and correctly rejects the confounded ones, maintaining the target FDR. This statistically-grounded framework provides a robust "causal compass" for experimental design. By identifying the sparse subset of parameters that truly control a material's properties, it offers a clear path to accelerating rational materials discovery and optimizing experimental resources.