invrs-gym 0.0.0

A collection of inverse design challenges

invrs-gym - A collection of inverse design challenges

v0.0.0

Overview

The invrs_gym package is an open-source gym containing a diverse set of photonic design challenges, which are relevant for a wide range of applications such as AR/VR, optical networking, LIDAR, and others.

Each of the challenges consists of a high-dimensional problem in which a physical structure (the photonic device) is optimized. The structure includes typically >10,000 degrees of freedom (DoF), generally including one or more arrays representing the structure or patterning of a layer, and may also include scalar variables representing e.g. layer thickness. In general, the DoF must satisfy certain constraints to be physical: thicknesses must be positive, and layer patterns must be manufacturable---they must not include features that are too small, or too closely spaced.

In general, we seek optimization techniques that reliably produce manufacturable, high-quality solutions and require reasonable compute resources. Among the techniques that could be applied are topology optimization, inverse design, and AI-guided design.

invrs_gym is intended to facilitate research on such methods within the jax ecosystem. It includes several challenges that have been used in previous works, so that researchers may directly compare their results to those of the literature. While some challenges are test problems (e.g. where the structure is two-dimensional, which is unphysical but allows fast simulation), others are actual problems that are relevant e.g. for quantum computing or 3D sensing.

Key concepts

The key types of the challenge are the Challenge and Component objects.

The Component represents the physical structure to be optimized, and has some intended excitation or operating condition (e.g. illumination with a particular wavelength from a particular direction). The Component includes methods to obtain initial parameters, and to compute the response of a component to the excitation.

Each Challenge has a Component as an attribute, and also has a target that can be used to determine whether particular parameters "solve" the challenge. The Challenge also provides functions to compute a scalar loss for use with gradient-based optimization, and additional metrics.

Example

# Select the challenge.
challenge = invrs_gym.challenges.ceviche_lightweight_waveguide_bend()

# Define loss function, which also returns auxilliary quantities.
def loss_fn(params):
    response, aux = challenge.component.response(params)
    loss = challenge.loss(response)
    distance = challenge.distance_to_target(response)
    metrics = challenge.metrics(response, params, aux)
    return loss, (response, distance, metrics, aux)

value_and_grad_fn = jax.value_and_grad(loss_fn, has_aux=True)

# Select an optimizer.
opt = invrs_opt.density_lbfgsb(beta=4)

# Generate initial parameters, and use these to initialize the optimizer state.
params = challenge.component.init(jax.random.PRNGKey(0))
state = opt.init(params)

# Carry out the optimization.
for i in range(steps):
    params = opt.params(state)
    (value, (response, distance, aux)), grad = value_and_grad_fn(params)
    state = opt.update(grad=grad, value=value, params=params, state=state)

With some plotting (see the example notebook), this code will produce the following waveguide bend:

Challenges

The current list of challenges is below. Check out the notebooks for ready-to-go examples of each.

• The ceviche challenges are jax-wrapped versions of the Ceviche Challenges open-sourced by Google, with defaults matching Inverse Design of Photonic Devices with Strict Foundry Fabrication Constraints by Schubert et al.
• The metagrating challenge is a re-implementation of the Metagrating3D problem using the fmmax simulator.
• The diffractive splitter challenge involves designing a non-paraxial diffractive beamsplitter useful for 3D sensing, as discussed in LightTrans documentation.
• The photon extractor challenge is based on Inverse-designed photon extractors for optically addressable defect qubits by Chakravarthi et al., and aims to create structures that increase photon extraction efficiency for quantum applications.

Install

pip install invrs_gym

Testing

Some tests are marked as slow and are skipped by default. To run these manually, use

pytest --runslow