Condition-gated graph neural networks with autoregressive training for plasma-etching profile prediction

Accurate prediction of plasma-etching profile evolution can support process exploration, but recursive surrogate predictions may accumulate shape errors. We present a condition-gated graph neural network (GNN) trained with dataset aggregation on a benchmark of 200 simulated Si etching processes in Cl2 plasma. Using the benchmark’s existing polar-coordinate representation, we treat each angular sample as a graph node and connect neighboring samples. Process conditions gate the features at each node, while a gated recurrent unit predicts profile evolution over time. On 29 common validation records, GNN-full achieves a mean intersection over union of 0.9904 for the final single step and 0.9561 for a 39-step rollout, compared with 0.9880 and 0.9487 for a retrained cascade recurrent neural network (CRNN). Corresponding mean radial-profile errors are 0.43 and 2.10 nm for GNN-full vs 0.59 and 2.78 nm for CRNN. Three-seed ablations identify nodewise conditioning and autoregressive training as the main contributors to long-horizon stability. A parameter-matched transformer retains similar performance, indicating that the angular-mixing mechanism is less decisive. Training on 50 selected records reduces training time approximately threefold while retaining most autoregressive accuracy. Both methods reproduce major process-condition trend signs, but CRNN has a lower vertical-depth error in the one-factor-at-a-time sweeps. GNN inference takes approximately 1.4 ms per step, with greater training and repeated-inference cost than CRNN. The results demonstrate a benchmark-specific accuracy–cost trade-off; manufacturing significance and prospective generalization require further validation.