Research / Featured research

Fluid-Body Interaction and Flow Control

Bodies and flows shape each other. We study how a deforming body sheds vorticity in order to move, and how a flow field can be prescribed in the first place: reinforcement learning finds the escape kinematics of larval fish under an energy budget, and a data-driven model turns a wall of fans into wind on demand.

Physical Review Fluids · 2021

Learning swimming escape patterns for larval fish under energy constraints

I. Mandralis, P. Weber, G. Novati, P. Koumoutsakos

How should a larval fish escape a predator when energy is limited? Reinforcement learning coupled to flow simulations rediscovers the classic C-start and also finds more economical escapes: covering the most distance on a fixed energy budget calls for short bursts of acceleration interleaved with gliding.

Paper

Vorticity shed by two escaping larval fish, above midline kinematics of a learned escape and a C-start — **Escape patterns, learned and classic.** The vortices shed by two escape manoeuvres (top), and the body midlines of an escape discovered by reinforcement learning beside the classic C-start (bottom).

Flow · 2026

Data-driven modelling for on-demand flow prescription in fan-array wind tunnels

A. Stefan-Zavala, I. Scherl, I. Mandralis, S. L. Brunton, M. Gharib

A wall of individually controlled fans can, in principle, produce any wind field, but turning a target flow into fan commands is a high-dimensional problem. From measurements alone we fit a surrogate model that maps fan speeds to time-averaged velocity profiles, invert it to prescribe flows on demand, and find the physics to be dominantly linear: a superposition of self-similar jets.

Paper

Schematic of a fan-array wind tunnel with fan-speed and velocity profiles, and the linear surrogate model — **Shaping wind with a wall of fans.** A profile of fan speeds (ii) produces jets that mix downstream into velocity profiles (iii), recorded by a sensor array (iv). A regularised linear map, A·r + b, predicts the flow from the fan speeds and, run in reverse, finds the fan speeds that produce a target profile.

In preparation · 2026

Robotic water jetting with an elastically loaded artificial mantle

joint work with J. Ro, P. H. Nguyen, M. Kovač

The Proteus project asks how a robot can propel itself by changing shape. An elastic artificial mantle stores energy as it is filled and releases it in a jet of water, an actuation principle that comes from the body itself rather than from a thruster. We are studying the fundamental questions behind it — how morphology, elastic storage and the surrounding fluid combine into useful thrust — with the goal of robots that sample and monitor lakes and other bodies of water.

More research Ground-Aerial Morphing Robotics · Learning-Based Control · Vision-Based Navigation and Manipulation