The SynLab works at the interface between data-driven technologies and societally relevant challenges. It develops advanced methods in mathematical optimization and machine learning to support decision-makers in sectors such as energy, health, and mobility. Over the previous two funding phases, the lab built the SCIP Optimization Suite into a powerful platform that makes high-performance optimization methods available across a wide range of applications. In the current phase, the focus is on bridging mathematical optimization and machine learning and on leveraging modern hardware, from GPU-accelerated solver components to resource-efficient deep learning.
Projects
Sustainable Machine Learning and Machine Learning for Sustainability
The project develops resource-efficient deep-learning methods — advanced pruning, compression, and data-pruning techniques that reduce the energy and compute cost of training and evaluating large neural networks while preserving accuracy, fairness, and robustness, and applies them to sustainability applications such as mapping global forest cover, biomass, and CO₂ concentrations from satellite data.
Next Generation Optimization Methods
The project develops new solution strategies for nonlinear and mixed-integer nonlinear optimization problems, with a focus on integrating first-order methods into the branch-and-bound framework. Topics include preprocessing tailored to first-order methods, generalized convexity within branch and bound, advanced heuristics that reduce the size of the search tree, and adaptive strategies that remove the need for expensive hyperparameter tuning.
High-Performance Optimization Software
The lab leads the development of the SCIP Optimization Suite (https://www.scipopt.org/). Its core, SCIP, is one of the fastest solvers for mixed-integer linear and nonlinear programming available in source code — free and open source under the Apache 2.0 license, and used in research, industry, and teaching worldwide. Current work focuses on accelerating the solver at its core: finding feasible solutions more efficiently, learning from infeasibilities through conflict analysis, GPU-accelerated first-order methods for large-scale problems, parallelization of core solver components, and bringing AI into the solving loop.
