Deep Reinforcement Learning-based AI agents, developed, trained, and deployed using Microsoft Project Bonsai tool chain, open new possibilities of process manufacturing improvements.

Coupled with the appropriate simulator to drive their training phase, those AI agents – aka Bonsai “brains”- can help improve processes such as the ones found in the production of chemicals, paper, or (textile) fibers. They can also help optimize most types of extrusion from food to metal and plastic.

Those AI agents can help reduce waste, startup time, specification drift, lines shut down time, and energy consumption.

To help customers test, on their own terms, how those use cases could be enabled with Project Bonsai, Microsoft now offers additional in-box demos. Customers can test and evaluate their specific use case before deciding to engage with a company such as Neal Analytics to build a custom AI agent for their process.

In the video below we show step-by-step how to use the plastic extrusion demo. Neal Analytics was commissioned by Microsoft to develop both the physics-based plastic extruder simulator as well as the Bonsai Inkling (the tool chain training programming language) code that trains the brain. This allows Neal Analytics AI experts to not only know in detail how to help any customer more effectively to adapt this demo to their specific situation, but also to understand the demo scope and limits.

This demo was built using the following assumptions. These are obviously simplifications of real-life use cases, but they allow for a realistic representation and demonstration of how such a plastic extrusion project would perform.

Plastic extrusion demo assumptions

• The simulator uses a physics-based approach modeling the fluid dynamics of PVC in a simplified extruder
• The extruder produces plastic rods of a given diameter with a target length defined by the user in the demo. The best representation is to think of those plastic tubes as plastic “spaghettis” as the die modeled is a simple circle.
• The real-life start-up process while the extruder warms up is ignored and the model assumes the plastic temperature in the extruder is the nominal one from the get-go
• Finally, the brain supports two optimization “concepts”:
  • A single goal one, “length optimization”, focuses on ensuring the rods’ length is within specifications
  • A dual goal one, “Yield optimization”, tries to minimize waste (rods that would be longer than the minimum needed) while ensuring the product remains within specifications  

How to test the demo by yourself?

The demo video below provides a step-by-step guide that describes how to create your own demo on your own Bonsai instance.

1. The demo can be found from Bonsai’s Azure site at https://preview.bons.ai
2. Ensure the “Beta” toggle is on
3. Select the extruder demo
4. Optional: change parameters such as target rod length, tolerance, training length in cycles, etc.
5. Train the first concept (it will take about 25-30 minutes)
   (You will be able to visualize during training how effective is the brain at ensuring (in success %) the concept goal(s) is/are reached.)
6. Train the second concept (same duration)
7. Train the selector
8. Export the brain to your IoT/Edge device or to the Azure cloud

Note: Each change in 4) will require 5) and 6) to be retrained.

Demo video 

This video provides a walkthrough of the steps described above (in accelerated form for the training steps).

If you have any question about the simulator, the curriculum or the overall feasibility of using this approach for your extrusion or other manufacturing process, please contact us.

Demo limits

This demo is not a full accelerator as, for real life use cases, several elements will need to be customized:

• The simulator is too simplistic to sufficiently represent an operating extruder behavior. The best approach will, in most cases, but to have the customer’s partner build a data-based AI simulator similar to what PepsiCo did for its extruder + oven simulator.
• The concepts are also simple ones and, depending on the customer’s priorities, new ones will have to be created by modifying the demo concept or creating new ones

However, the demo is realistic enough that it can help customers virtually PoC whether Project Bonsai could be an applicable solution to their process optimization challenges.

Learn more:

• Autonomous Systems: Simulations
• Source code for the extruder on Microsoft GitHub repo: https://github.com/microsoft/microsoft-bonsai-api/tree/main/Python/samples/plastic-extrusion
• Step-by-step guide: Getting started with Project Bonsai
• Production Yield Optimization