Adaptive Learning, Searching and Optimization
Adaptive Learning, Searching and Optimization is MatterSpace’s energy-based research direction. It studies how to search very large design spaces for candidates that meet hard requirements, and how to spend expensive computation where it is most likely to pay off.
Developed by Faruk Guney.


Our thesis: energy-based approaches will matter more and more.
We believe energy-based approaches, including energy-based models (EBMs), will play an increasingly important role in how AI samples and evaluates candidate designs for science, research, and engineering.
This is the research thesis our work is built around: a view of where the field is heading, which we test on real design problems.
Requirements shape the search
Hard requirements and flexible preferences guide which candidates are considered from the start, rather than sorting finished results.
Physics informs every candidate
Physical plausibility shapes how candidate designs are proposed and how they are judged.
Computation where it counts
Faster evaluations come first, and heavier physics-based calculations go to the candidates that merit deeper validation.
Design spaces, today and next.
In materials, a design space spans many dimensions at once. Chip design is the next area we plan to research.
Structure
How the atoms of a material are arranged.
Composition
Which elements a material contains, and in what proportions.
Processing conditions
How a material is made, such as the temperatures and pressures it goes through.
Chip design, next
A future research area: chip design candidates that meet targets such as performance, power consumption and area.
Two loops, seen from your side.
The search works in two loops, and each has a job you can see.
The inner loop
Proposes candidates and gives each one a first evaluation.
The outer loop
Uses those results to set the direction of the search and to help decide which candidates go on to deeper validation.
MS Lattice applies this research to materials today.
MS Lattice is physics-based inverse design for materials discovery and optimization. Your requirements and preferences guide candidate generation from the start, and computation goes to promising materials worth validating. Computational promise, computational validation and experimental validation are separate steps: MS Lattice delivers the first two, and experimental validation happens in a lab, your own or through Norck.
Developed by Faruk Guney.
Bring us a design problem.
Tell us what a material must do and the limits it must respect, and we will show you how MS Lattice would approach it.

