Autoencoder Fragrance Composition Generation With Viability Constraints

The patent replaces traditional statistical optimization techniques with neural network-based machine learning systems. The neural network learns complex patterns and relationships in ingredient compositions through training on large datasets, enabling it to handle high-dimensional ingredient spaces (thousands of ingredients) that statistical methods cannot manage. The network architecture includes embedding layers that transform ingredient identifiers into continuous vector representations, allowing efficient processing of large ingredient dimensions.

The patent uses dynamic generative models (GANs, VAEs) that can adaptively generate new compositions based on learned patterns rather than fixed rules. The neural networks maintain control through learned constraints and objectives while generating indeterministic new compositions. The system dynamically adjusts generation parameters and explores the composition space creatively while respecting formulation constraints.

The patent incorporates constraint satisfaction and viability filtering directly into the neural network training process and generation architecture. The networks are trained to inherently produce viable compositions by learning from valid example compositions and incorporating domain knowledge constraints. This preliminary action during training eliminates the need for extensive postprocessing filtering, as the generated compositions are already more likely to be viable.

The patent merges individual ingredient olfactory property predictions with composition-level analysis. The neural network models learn not only individual ingredient characteristics but also how ingredients interact and combine to produce overall composition effects. The system integrates unitary molecule properties with emergent composition properties, accounting for non-linear interactions between ingredients and solvents in the composition.