Machine-learning-driven prediction and multi-objective optimization of Ta₂O₅ film performance in ion-assisted electron beam evaporation

In high-energy laser systems, the performance of tantalum pentoxide (Ta₂O₅) thin films prepared by ion-beam assisted electron beam evaporation (IAD) is extremely sensitive to deposition parameters. Due to complex physical coupling mechanisms, slight variations in any parameter can trigger micro-defects, stoichiometric deviations, and stress fluctuation, making it challenging to identify the optimal process parameter window in engineering applications for achieving high refractive index (n), low absorption (A), and low residual stress (σ). To address this issue, deposition experiments were conducted to obtain data on process parameters and film performance, which were utilized to establish the initial population for the Non-dominated Sorting Genetic Algorithm II (NSGA-II). Subsequently, machine learning was employed to construct prediction models for n, A, and σ. A voting-based ensemble model combining Random Forest (RF), Extremely Randomized Trees (ExtraTrees), and Extreme Gradient Boosting (XGBoost) was adopted, where RF improves prediction stability under small-sample conditions, ExtraTrees enhances robustness against experimental noise through increased randomness, and XGBoost captures complex nonlinear process-property relationships. Experimental validation was further performed using nine representative Pareto-optimal solutions. The average relative errors for n, A, and σ were 0.07%, 10.47%, and 1.68%, respectively, while the comparatively higher relative error for A was mainly associated with its low absolute values and greater sensitivity to microscopic defects and measurement fluctuations. Furthermore, considering diverse optical requirements, solutions prioritizing n, A and σ were obtained via the fast non-dominated sorting algorithm combined with an elite strategy. The optimized process windows were all centered near I = 850 mA, while the preferred V, F_O2, and F_Ar ranges varied with the target priority: 600–700 V, 50–60 sccm, and 5–6 sccm for n; 300–600 V, 55–60 sccm, and 5–10 sccm for A; and 300–400 V or 790–810 V, 50–60 sccm, and 5–10 sccm for σ. Analysis of the Pareto frontier revealed that, within the studied parameter range, n and σ exhibited a weak correlation, whereas clear trade-offs existed between n and A and between A and σ. Representative XPS, XRD, and AFM characterizations of selected Pareto-favorable and comparatively poor-performing samples further showed that the performance deterioration under excessive ion-beam input was mainly associated with defect-related oxygen states and surface roughening, rather than crystallization-driven changes. These results demonstrate that the proposed surrogate-assisted optimization framework can resolve complex parameter trade-offs and provide technical support for the high-precision IAD of optical thin films with high n, low A, and low σ.