Machine-learning techniques are highly advantageous for automation of a manufacturing process, since they facilitate prediction of the process parameters and product properties. However, the need for process-specific prior knowledge, for elaboration of complex analytical models, and for collection of a comprehensive training dataset notably limits their integration into the real-world applications. We pioneered and studied usage of a streamlined non-analytical approach to predict and optimize process parameters, radically adapted to the conditions of resource and knowledge constraints. The approach was employed for laser-induced graphene, which is an emerging flexible-electronics fabrication technique. The fabrication settings were successfully predicted and controlled by a blackbox neural network from the desired properties of the device, despite a small amount of moderate-quality training data. To prove feasibility of the concept, we designed and manufactured a functional electronic circuit. The proposed procedure is applicable for a broad range of functional materials and fabrication methods.

The datasets encompas collected training, validation, and test data as well as related mesuared LIG properties. Five dataset include (A) 500 training lines, (B) 100 validation lines and (C) 13 validation rectangles, (D) 50 twofold-control lines, and (E) 10 printed lines used for the demonstrator-circuitry characterization. For each collected sample, laser-beam parameters, and corresponding product (LIG) properties are presented. "Set" indicates input data: grid-distributed in the dataset A, randomly generated in the datasets B-D, and a single calculated value in the dataset E. "Predicted" stands for the neural-network predicted output data. "Measured" represents the actual, physically and optically measured sample properties.