AI Driven Design Approach - 2020

Details

Title : AI Driven Design Approach Author(s): Machine Learning Center at Georgia Tech Link(s) : https://www.youtube.com/watch?v=eitWsEM2ljk

Rough Notes

The design process is complicated, communication between the customer, the project manager, the engineers and the technicians can result in a design that was not what the customer initially wanted (e.g. shown by speaker here).

Outline of talk alongside notes:

  • Introduction: Generative Engineering for Design The current design process starts from defining/formulating the problem, then creating the solution resulting in seed designs which come from the knowledge (including tacit) from engineers, which are then evaluated to check whether they meet all the requirements. The last two steps are repeated until we reach the requirements and/or run out of budget. There is no one person who knows about everything.

    We want to capture the engineers knowledge (including tacit) and achieve design discovery efficiently.

    The big picture involes capturing relevant information and storing them, and discovering the relevant information, and making relevant discoveries in them.

  • Methods: Knowledge Graph This makes it easy to extend and merge new knowledge. E.g. the speaker's group has a knowledge graph on subtractive manufacturing, which contained information about the different resources, what kind of capabilities they have, what kind of operations they can do, what are the skills they can provide, etc to come up with the most optimal way of manufacturing something like a car. To add knowledge about additive manufacturing, they just need to add a subgraph with the same similar scheme and make the relevant connections, and the algorithms did not need to change. Now the knowledge graph can be utilized to do hybrid manufacturing, how to make a part using additive and subtractie manufacturing.

    Knowledge graphs also make it easier to see high level information, by making subgraphs a node, seeing the graph at a different scale.

    These graphs are generated via different distillers that take XML, CSV, JSON etc with algorithms independent of the problem at hand.

    New knowledge is considered to be things like:

    • New constraints/rules derived from past/new data.
    • Generation of new relationships between entities.
    • Generation of new entities.
    • Combining existing constraint/rules.
    • New solutions.
  • Methods: Forward and Inverse Models Some e.g. for forward models include using ResNets as surrogate models to predict gain and frequency given design parameters for a power circuit board. For inverse models, mixture density networks, GANs etc are used to estimate the input given the output.
  • Methods: ML driven Optimization This involves combining forward and inverse models and use them to perform some optimization objective. Traditionally, this starts from "cold" initial designs/seed designs, which are then analyzed via simulations until the design reaches a stopping criteria. We would want to invert this process, start from a design target, and generate "hot" initial design, and do small changes to them until we reach a goal we are satisfied with.
  • Methods: Dependency and Causal Analysis Early in the design we want to know which parameters which are most responsible for the design being in a certain state. This can be done via a partial correlation based approach to do influence analysis, i.e. given a network it finds out the parameters which are most influential for the output, and which nodes are most influenced by other parameters.
  • Putting it together: Turbine Design Here the expert gives requirements, an inverse model takes these, generates new seeds and design parameters. The design parameters go to a forward model, whose outputs alongside the initial seeds go to a ML based optimization procedure, which can be optionally fed into a traditional optimization procedure. The resulting design is then sent to the expert and the procedure iterates. In practice, the ML solution alone was not as good as the design suggested by existing software (HEEDS), however using the ML generated design as the first starting point for it gave better results than using existing software alone.
  • Putting it together: BOP Generation One component here includes a NLP based distiller which takes unstructured data from engineers, and classifies each word into a product process/resource, and then learn which steps are creating which manufacturing feature. E.g. to create feature A,B, these are the different ways of manufacturing it.
  • Research issues to be addressed:
    • Interactive design space exploration. The user knows the design they want but might not be able to quantify mathematical equations. How can we learn the tacit knowledge, i.e. the bias towards a certain design.
    • Design space characterization. From the data can we identify good, and bad designs and those which have specific features.
    • Automatic constraints/design rules generation.
    • Automatic formulations of objectives.
    • Automatic surrogate generation.
    • Capture and utilization of human (design engineer's) bias.
    • Design for designer's intent.
    • Parametric mapping across scales.
    • Automatic knowledge capturing from requirement documents.
    • Time evolving knowledge for design version control.
    • Transfer learning cross-domain or cross-physics design problems.
    • Which ML algorithms to use, and why etc.

Emacs 29.4 (Org mode 9.6.15)