Accelerating Rational Crystal Habit Design via Interpretable Mechanism-Guided Machine Learning Framework

The pharmaceutical crystal habit significantly influences
downstream
processing and product quality. However, traditional experimental
screening methods rely heavily on empirical knowledge and trial-and-error
experimentation, demanding substantial resources and time. Here we
present an interpretable machine learning framework integrating molecular
descriptors with quantum chemical (QC) or thermodynamic (HSP, PC-SAFT)
parameters to predict crystal habits. Using a curated database of
418 entries covering 153 APIs and 41 solvents, the XGBoost model demonstrated
superior performance among seven algorithms. Models augmented with
QC or thermodynamic descriptors significantly improved prediction,
achieving AUC > 0.88. SHAP analysis revealed dipole moments and
solubility
parameter differences as key determinants of crystal habit, enhancing
interpretability. External validation on six compounds in seven solvents
confirmed that models with additional mechanistic descriptors consistently
outperformed the molecular-only baseline, with the Molecular+QC model
showing the best overall performance. This proposal paves the way
to the rational design of crystal engineering strategies for habit
prediction and control, and offers a general framework that can readily
be extended to other crystallographic applications, as well as broader
domains, including drug discovery and materials science.