A morphology slot for literature-trained solar-cell models: the 3HT decamer:JSM5 blend's blended PCE prediction moves from 4.0% to 4.5%
Synopsis
Adding a morphology input channel to an organic-photovoltaic prediction pipeline that previously held only molecular descriptors raises the blended PCE prediction for the 3HT decamer:JSM5 blend from about 4.0% to about 4.5%, and the PCNN alone from about 2.4% to about 3.7%. The channel is a set of latent placeholder slots learned from device-target residuals and filled at prediction time with molecular-dynamics morphology descriptors; because the latents are target-derived, the shift reflects the pipeline's response to MD inputs and cannot be attributed to a physical morphology effect.
Interpretation
Organic-photovoltaic device prediction can take morphology descriptors computed by molecular dynamics, not only molecular electronic structure. The training data came from 2,080 literature devices carrying donor and acceptor HOMO, LUMO, optical gap and device architecture, with no morphology descriptors. For the 3HT decamer:JSM5 blend, adding MD descriptors moves the blended PCE prediction from 3.998% to 4.478% and the PCNN from 2.385% to 3.685%.
Morphology enters through latent slots learned from target residuals rather than as a direct training feature. The training set has no morphology columns, so a random forest is fit per target on molecular features, and five-fold out-of-fold residuals are standardized, reduced by PCA, and rank-mapped onto four physical ranges: pi-stacking distance, donor-acceptor distance, orientational order and crystalline fraction. Each training device receives latent values from out-of-fold predictions that exclude its own targets, while validation and test devices use a mapper fit on the training partition only, avoiding circularity.
The MD descriptors bring the two model families closer together. Without MD descriptors, the baseline ensemble and the PCNN differ by 3.65 percentage points in predicted PCE (6.038% versus 2.385%). With them, the gap narrows to 1.60 percentage points (5.288% versus 3.685%), although the two move in opposite directions for Jsc and PCE.
Perspective
The demonstrated setting is the single 3HT decamer:JSM5 donor-acceptor combination, trained on 2,080 literature devices after consistency filtering and unit standardization and tested on a grouped held-out split of about 310 devices. The authors state that the influence of MD descriptors on predicted performance will require validation and optimization against independently obtained morphology information, together with a sensitivity analysis of the latent-slot construction.
The authors state that the latents are target-derived and can only absorb residual variation, so their inclusion should not be read as solving the missing morphology physics; predictions regress toward the mean at the extremes and the nominal 95% intervals under-cover in the tails; and the MD simulation omitted side chains on both donor and acceptor, which the authors say needs recalculation.