Computational biophysics

Understanding biological function.
Engineering what comes next.

We model how biological systems move, sense, and respond. Then we ask whether a different material or mechanism could produce the same behavior under the same constraints.

MCF-7 signaling · SIML CUDA vs. Triton · One RTX 306034.51% more simulated cells under a 6 GB limit →
Conceptual experimentTwo ways to move a load
Biological reference

Tendon-driven finger

Proposed candidate

Electrically driven strip

Apply the same load. Compare force, travel, and response time. Physical testing of a candidate is planned.

Artificial muscle study

Can a different physical system reproduce the same biological function?

Start with what the system does. A candidate can use a different structure or mechanism if it meets the same measurable requirements. These are three questions we can put to a test.

Vision

Sensing and tracking

Could a different sensor and controller recover a target after it passes behind an obstacle?

From a reference behavior to a physical test.

  1. Biological reference

    Choose a behavior and the conditions under which it is observed.

  2. Functional specification

    Define the quantities, units, and tolerances a candidate must match.

  3. Simulation

    Choose a solver for the physics and check it against a reference.

  4. Candidate exploration

    Compare modeled responses and identify a discriminating test.

  5. Physical validation

    Test a specimen under the same conditions. This stage is planned for materials.

Explore the scientific method →

Demonstrated results

What we measured on the GPU.

For a research team with limited GPU memory, the first question is how much of a model will fit. A separate question is how quickly a particular kernel updates its state. These studies answer those questions on two different workloads.

Biological Simulation

GPU capacity for cellular signaling

A GPU can run out of memory before a model reaches the population a researcher wants to study. On one RTX 3060, SIML CUDA advanced 8,100,736 interacting MCF-7 signaling patches through six simulated seconds. Triton advanced 6,022,368 under the same 6 GB budget and accuracy requirements.

Computational Systems

GPU updates for binary lattices

On one RTX 3060, SIML updated 256 small binary lattices 12.08× as fast as a separately tuned UInt8 CUDA control once the data were on the GPU. A separate capacity test kept one connected lattice of 84.26 billion binary variables updating continuously for ten minutes.

Compare published measurements

Study
34.51%more simulated cells under a 6 GB limit

On one RTX 3060, both populations completed six simulated seconds of the MCF-7 signaling benchmark and passed its accuracy checks.

Simulated cells
SIML CUDA
8,100,736
Triton
6,022,368
Discuss a technical evaluation →

Materials research

What would it take to match a muscle?

Start with the force and strain a muscle produces under a given load. Measure how quickly it contracts and relaxes. A candidate ionic polymer would also need testing for electrical energy use, hysteresis, fatigue, and stability. Those measurements would tell us which properties it can reproduce.

See the artificial muscle milestones →