2.11. Criteria for a Successful Prediction

Although the computational efficiency of the AMOEBA force field is essential for the alchemical polymorph search, its accuracy for lattice energies compared with modern periodic DFT methods that include a dispersion correction has not been thoroughly explored. Experimental polymorphs for compound XXIII were obtained from the CCDC and compared to the final predicted polymorphs from the “precise” DFT minimization procedure using the PAC algorithm with molecular clusters of 20 molecules (rmsd₂₀). An rmsd₂₀ of 1.0 Å or less between the predicted polymorph and the experiment was considered a success. For AMOEBA-minimized structures with an rmsd₂₀ of 1.5 Å or less compared to the experiment, downstream DFT-D minimization resulted in rmsd₂₀ values below 1.0 Å. This observation is based on the data summarized in Table S1, which shows that predicted structures starting from an AMOEBA rmsd₂₀ to experiment of ≤1.5 Å and were promoted to DFT minimization ultimately achieved an rmsd₂₀ to experiment of ≤1.0 Å. The 1.0 Å cutoff for the final similarity ranking was chosen based on crystal structure similarity metrics used previously.¹¹ These cutoffs can be adjusted depending on the system (e.g., larger APIs where more variance is observed but maintain similar crystal packing).