Scientific ML Studio

About us

Understand. Experiment. Build.

Most machine-learning teaching starts from data and hopes physics turns up later. Scientific ML Studio goes the other way round: you already have the governing equation, the domain and the boundary conditions, and the network's job is to satisfy them.

Start reading Open the Lab

What lives here

One idea, three places to meet it

  1. Understand

    Learn is an open textbook: chapters on the left, the mathematics set properly, your place remembered between visits. Beside it sit short video guides, starting with a walk through every block in the Lab. The reading stays free.

    Open the textbook

  2. Experiment

    Lab is the workbench. Wire a problem together on a canvas (domain, equation, conditions, network) and change anything to see what happens. It runs entirely in your browser, and nothing you build is sent anywhere.

    Open the Lab

  3. Build

    Every notebook turns into a PyTorch script you can download and run on your own machine. The Lab is where you sketch the problem; the script is what you take away, read and defend.

    Watch the guides

The idea

Every constraint pulls toward zero

A physics-informed network is not shown the answer. It is shown how wrong it currently is: the residual of the equation, the mismatch at the boundary, the mismatch at the start. Training is the process of pushing all of those down together.

That is why the picture at the top of this page has many lines pulling into one place. Many separate requirements, one solution that satisfies them all.

Loss terms falling together during training (schematic) Three curves on a logarithmic scale, the PDE residual, the boundary term and the initial-condition term. All start near 1 and fall by three to four orders of magnitude as training proceeds. This is an illustration, not measured data. 110−110−210−310−4 0%25%50%75%100% Training progress Loss (log scale) PDE residualBoundaryInitial
  • PDE residual
  • Boundary
  • Initial
Schematic. An illustration of the idea, not measured data.

How the Lab works

Blocks in, PyTorch out

Each block is one decision about the problem. Connect them and the Lab checks the wiring, then writes the script. The same generator runs in your browser, so what you see is what you get.

  1. DOMDomain
  2. PDEEquation
  3. BCBoundary
  4. NETNetwork
  5. LOSSLoss
  6. OPTOptimiser
  7. RUNTrain
Try it in the Lab Watch the Lab guide

How we build it

Three promises

  • It runs on your machine

    The Lab executes Python in your browser. There is no account to make and no server holding your work.

  • Your notebooks are yours

    They are stored on your device, and you can download a copy at any time to keep it or move it elsewhere.

  • The code is the same code

    The script the site writes is identical, character for character, to what the reference implementation produces from the same graph. We test that on every change.

Who is behind it

A small team, working in the field

Solid mechanics, micromechanics and physics-informed neural networks.

  • Portrait of Dr. Fallah

    Dr. Fallah

    Founder of Scientific ML Studio

    Dr. Ali Fallah is an Assistant Professor of Automotive Engineering at Atılım University in Ankara, Turkey, specializing in computational solid mechanics, smart structures, additive manufacturing, and multiscale modeling. His research also explores physics-informed machine learning, including PINNs and neural operators, for advanced engineering analysis and simulation.

  • Portrait of the second team member

    Sherwyn

    Their role

    One or two sentences on what they do and what they are known for.

Read it. Try it. Tell us what you think.

Start reading Contact us