Model Accelerated Phase Space Search
MAPSS is our proprietary method to substantially reduce the amount of computations needed to find solutions. MAPSS uses machine learning to guide the search for potential solutions. This solves one of the top complaints optical design software users have about existing commercial solutions. MAPSS modules can be combined to extend their capabilities without the need for generating new models for each potential problem.
The MAPSS approach
Conventional solution techniques pair a physics calculation software with an optimizer algorithm that iteratively alters design degrees of freedom (DOFs), runs performance simulations and determines if the change has improved the performance of the design. For example, when developing a camera lens, a conventional optimizer would adjust the values of radii of curvature for different lenses with the goal of finding the best resolution, while maintaining constraints on the overall size of the assembly.
Large phase spaces, and wasted computations
When there are many degrees of freedom, the optimizer algorithm has a much larger phase space it must hunt through before stumbling upon a change which improves performance. For each iteration, the optimizer must run a full performance simulation. Since it is improbable to immediately select the optimal degree of freedom for iteration, most changes either decrease or have no improvement on performance. This wastes significant computational resources. Because the optimizer is slowly hunting for design improvements, the output of the process is not entirely deterministic: It’s possible to miss best possible solutions because the optimizer goes off in the wrong direction.
Using machine learning
MAPSS reduces the computational waste in two ways. First, it provides the optimizer it’s best estimate of the global optimum for a solution. Second, for each iteration, it provides the optimizer guidance about which DOFs to alter in order to improve the result. For designs with many DOFs, this substantially improves the efficiency of the search process and reduces the likelihood of missing acceptable solutions.
Benefits
Compute-cost reduction
MAPSS not only improves the efficiency of solution searches, it reduces the variance in the compute resources used for similar classes of problems. This allows forecasting of compute expenses based on usage and selection of cost-optimized compute nodes in the cloud.
Additionally, MAPSS eliminates the risk of wasted significant compute resources searching for an infeasible performance request. For traditional optimizers, the only way to determine a request is infeasible is to run iteration after iteration until the compute budget is consumed. MAPSS flags the request as infeasible early in the design search process, and can provide recommendations for how to make the request more realistic.
Design breadth
By reducing compute costs along with their variance, it is possible to run parallel design searches with different numbers of components with different constraints. For example, in an imaging application it is possible to run design optimizations in parallel on Cooke triplet design concepts, as well as Clarke double Gauss, Zeiss Planar, etc. enabling the customer to understand what is possible for each design concept.
Catalog components
A particularly challenging design space to search through is that of catalog optical components. The challenge stems from the discontinuous nature of the degrees of freedom. For example, changing the focal length of a lens can result in a simultaneous change in the clear aperture based on how the component search is ordered. The use of a machine learning model reduces the risk that the ordering of components will have any impact on the final design result.