Data-Driven Topography and Roughness Analysis of Plasma Deposition Processes for Diamond-Like Carbon Coatings

Surface roughness regulates the biological and electrochemical properties of diamond-like carbon (DLC) coatings, depending on ion kinetics from surface area per unit volume. The common DLC manufacturing methods lack control over selective deposition and require resource-enriched systematic study of multivariable studies to achieve the desired level of surface roughness for targeted applications. This work develops and validates a multi-stage data-driven framework combining Gradient Boosting (GB), Multi-Layer Perceptron (MLP), and a hybrid GB–MLP approach to predict diamond-like carbon (DLC) roughness from 1,000 curated datapoints covering nine deposition parameters. Gradient Boosting achieved the highest accuracy for parameters with sharp transitions (e.g., voltage, thickness), outperforming other models in 52–92% of cases, while MLP provided superior performance for smooth nonlinear parameters (up to 80% accuracy). A hybrid GB–MLP approach reduced prediction variance and produced the most stable multivariable roughness predictions. A physics-informed Convolutional Neural Network (CNN) is trained to generate surface height-maps from given roughness values, achieving greater than 95% reconstruction fidelity with Average roughness (Ra) errors below 5x10 -5 μm. This benchmark is set by examining the difference between actual Ra values obtained from Atomic Force Microscopy (AFM) experiments and a digital heights map surface image reconstructed with integrating of Ra governing equation. The combined numerical–generative pipeline reproduces known trends in DLC deposition (e.g., minimum roughness at optimal bias voltage, thickness effects, and ion-energy. Finally, an analytical model linking Kurtosis (Sku) to deposition parameters is proposed for functional performance evaluation and digital-twin integration. To our knowledge, this is among the first DLC studies to integrate (i) multivariable ML prediction, (ii) physics-informed generative microtopography with quantified image-level error, and (iii) a functional analytical model (Sku) in a single pipeline validated against literature-curated experimental data.