AI and Machine Learning Algorithms for Progressive Freeform Lens Design

Artificial intelligence (AI) is transforming multiple industries, and ophthalmic optics may represent a new field of application. In progressive lens engineering, machine learning algorithms could provide new methods for optimizing complex optical parameters and improving the final visual experience of wearers. The study by Gaetano Volpe and Pasquale Fanelli examines the main AI learning strategies and proposes their potential use in progressive freeform lens design.

Machine learning is a subset of artificial intelligence that enables machines to learn how to perform specific tasks through algorithms and data. Among its main approaches are supervised learning, unsupervised learning, reinforcement learning and multitask learning. Supervised learning uses labelled datasets to establish correlations between inputs and known outputs, allowing the system to make predictions when new data is introduced.

By contrast, unsupervised learning searches for hidden patterns within unlabelled datasets, while reinforcement learning improves decisions through trial and error, using rewards and penalties. Multitask learning allows a single model to perform several tasks simultaneously, exploiting shared information to improve overall performance.

A central role is played by Artificial Neural Networks (ANNs). Inspired by biological neural networks, these systems process information through interconnected nodes whose outputs are determined by different weights. When neural networks contain numerous hierarchical layers and large datasets, they become deep neural networks, or deep learningsystems, capable of handling highly complex tasks.

The most innovative perspective concerns their application to progressive ophthalmic lenses. Designing a high-quality progressive lens requires optimization of the optical surface to minimize optical aberrations while maximizing visual performance. Machine learning, particularly supervised learning, could model the relationship between lens parameters and the wearer’s perception.

Customer satisfaction questionnaires could become valuable training data. Information concerning optical power, lens geometry, mounting and position-of-wear parameters could form the neural network’s input, while wearer satisfaction scores would represent its output. As illustrated by the neural-network diagram on page 3, the system could progressively modify design parameters to identify configurations associated with higher satisfaction scores.

Continuous feedback could therefore allow the network to keep learning and potentially improve both predictions and progressive lens optimization. The authors stress that, at the time of their study, scientific literature did not yet contain publications applying these technologies directly to freeform lens design; their proposal is therefore a potential AI application scenario, rather than an already validated clinical method.

Download MAFO 04-23