AI-Driven control of large potassium dihydrogen predictive modeling and assisted optimization of DKDP crystal growth using gradient boosted regression trees and crystal height features

This work develops a machine‑learning assisted approach to model and optimize the growth of deuterated potassium dihydrogen phosphate (DKDP) crystals using industrial‑scale data. Conventional models based only on temperature and supersaturation fail to capture multi‑field coupling and size effects in large crystals. Here, crystal height is introduced as an additional physical feature and combined with a Gradient Boosted Regression Trees (GBRT) algorithm to predict growth rate. On 15,881 daily records from long‑term DKDP runs, adding crystal height increases the test‑set R² from 0.4916 to 0.8427 and significantly reduces MSE, RMSE and MAE. Feature‑importance analysis shows that temperature, saturation temperature and crystal height contribute 22.39%, 42.58% and 35.03%, highlighting the key role of height in size‑dependent heat‑ and mass‑transfer. The trained GBRT model enables real‑time prediction and anomaly detection, where large prediction deviations are associated with “hair” inclusions and lateral‑expansion defects. This grey‑box strategy offers a practical route to intelligent monitoring and assisted optimization of large DKDP crystal growth.