Quantum chemical calculations using density functional theory (DFT) are now commonplace in materials research. There are several publicly available materials databases (e.g., Materials Project) that house DFT calculation results for millions of inorganic crystals, providing an excellent starting point for finding new materials with interesting properties. High-throughput computational screening of materials using DFT is now being augmented and accelerated using machine learning (ML). In this talk, I’ll discuss some shortcomings in the “early days” of ML-driven materials discovery (circa 2020) before transitioning to state-of-the-art approaches. Modern approaches leverage universal machine learning interatomic potentials (“foundation” potentials) along with generative artificial intelligence (AI) to rapidly search materials space. From the standpoint of materials discovery, the first challenge is efficiently identifying new inorganic crystals that are thermodynamically stable. Once these materials are identified on the computer, the next step is to make them in the lab. Unsurprisingly, ML models are now emerging with the goal of predicting which hypothetical materials are likely to be synthesizable. While these developments are exciting, a pervasive challenge for any ML problem is to anticipate the performance of models outside the curated experiments used for initial training and evaluation. This talk will detail my group’s efforts to systematically assess the quality of generative models and synthesis prediction models through the lens of first-principles thermodynamics.