Fig. 3: Model construction, evaluation, and deployment.

a A simplified architectural diagram of the model. The preprocessed data is processed by a 3D ResNet 50 model, which comprises both residual and fully connected structures and outputs classification results. Conv3D is a 3D convolution layer. BatchNorm3D is a 3D batch normalization layer. ReLU represents the rectified linear unit activation function. Training performance of (b) “QDs model”, and (c) “density model”. Train: training. Valid: validation. Acc: accuracy. Loss: loss. The Loss and Acc curves are derived from results during model training. The Trend and Change curves are simple polynomial fit and the first derivative of the Loss curve, respectively. (d) A control logic diagram for the model deployment. After opening the shutter, the substrate temperature remains unchanged if the label output from the density model aligns with the preset target before quantum dot (QD) formation. If the output from the density model exceeds the target density, the substrate temperature increases; otherwise, it decreases. Once the QDs model recognizes QD formation, the shuttle will remain open until the density model’s output matches the target. At this point, it closes the shutter to complete the growth process. Source data are provided as a Source Data file.