Uncertainty-Aware Chemical Analyses via Monte Carlo Deep Model Inference

• Leandro Pereira Bernardo UFF
• Andre Luiz da S. Pereira UFF
• David Neves Dias UFF
• Helida Vieira UFRJ
• Renato Malbar UFRJ
• Aline Machado de Azevedo Novaes CENPES
• Rafaella Magliano Balbi de Faria CENPES
• Rogerio Mesquita de Carvalho CENPES
• Maria Luisa Aleixo Gonçalves UFRJ
• Ana Mehl UFRJ
• Leandro A. F. Fernandes UFF

Resumo

We present a computer vision-based solution for estimating Chloride concentration and pH in water samples. The proposed chemical methodology produces samples with visually distinguishable reactions. Images are captured in a controlled lighting environment, and preprocessing techniques are applied to segment the Region of Interest (ROI). Each ROI is processed by a Convolutional Neural Network (CNN), which extracts highlevel descriptors. From each input sample, a total of n×n CNN-based descriptors (one per pixel in the spatially aligned feature map) yield a high-dimensional representation of local sample characteristics. They are input into a Neural Network model. After training, the model can infer a predictive distribution for each input image. This allows for the computation of descriptive statistics that characterize the associated uncertainty of the model's predictions. Our approach is inspired by the Monte Carlo method: descriptors are treated as independent observations drawn from a latent feature distribution. Although they are deterministically extracted from activation maps, they result from random visual outcomes of the same underlying analytical parameter. Conversely, they are not independent. The proposed approach can be applied to several tasks, including quality assurance, process monitoring, and compliance verification.

Palavras-chave: Analytical models, Visualization, Monte Carlo methods, Uncertainty, Computational modeling, Predictive models, Feature extraction, Computational efficiency, Convolutional neural networks, Chemicals