Exploring celecoxib polymorph landscape using AIMNet2 machine learning interatomic potential

The crystal form a drug adopts can change everything from how it dissolves to whether it works in the clinic, yet predicting which polymorphs a flexible molecule will produce remains one of the most stubborn problems in pharmaceutical science. Competing forms typically differ in energy by less than 2 kJ mol^-1, a precision that quantum chemistry can reach only at forbidding cost. Here we deploy AIMNet2, a machine-learned interatomic potential refined by active learning on cluster reference data, to map the polymorphic landscape of celecoxib, a widely prescribed COX-2 inhibitor whose form I exhibits record-breaking elastic flexibility. A GPU-accelerated workflow generates and ranks hundreds of thousands of candidate structures at near-quantum accuracy, recovers the experimental ordering of forms I, II, and III with sub-Ångström geometric fidelity, and identifies two low-energy candidate structures within 4 kJ mol^-1 of the most stable known polymorph. Hybrid-DFT calculations yield a similar low-energy landscape in which multiple polymorphs remain thermodynamically competitive. Finite-temperature analyses further expose the limits of static-lattice models for ultra-soft crystals such as form I. Beyond celecoxib, the framework supplies physically motivated targets for experimental polymorph screening and a transferable strategy for crystal-structure prediction across flexible drug molecules.