CHANNELS

Aerospace

Automotive

Automated & Autonomous Vehicles

Battery & Electrification Technology

Defense

Drone TV

Electronics

Manufacturing & Prototyping

Materials

Medical

Motion Control

RF & Microwave Technology

Robotics & Automation

Sensors & IoT

Test & Measurement

Aerospace
Automotive
Automated & Autonomous Vehicles
Battery & Electrification Technology
Defense
Drone TV
Electronics
Manufacturing & Prototyping
Materials
Medical
Motion Control
RF & Microwave Technology
Robotics & Automation
Sensors & IoT
Test & Measurement
Aerospace & Defense
Search Results

DarkStar Inverse Design Framework: Design from the End Result

The DarkStar Inverse Design Framework helps researchers discover complex, optimized science and engineering design solutions by working in a revolutionary new way. It blends AI, machine learning, and advanced hydrodynamics simulations to optimize science and engineering solutions starting from the final state.

Topics:
Artificial Intelligence Augmented/Virtual Reality Automotive Predictive Models and Predictive Analytics

More From SAE Media Group

    Transcript

    00:00:00 to test their designs scientists rely on time dependent physical simulations using powerful supercomputers this however is no simple task designing Machinery from simulations requires a great deal of trial and error forcing scientists and Engineers to iterate but what if they didn't have to test their designs moving forward in time what if they could start

    00:00:27 at the desired end result and run the physics simul ation backwards to see what caused it this is a daunting challenge but Lawrence Livermore National Laboratory has found an Innovative solution a multi-disciplinary team of Livermore scientists and Engineers embarked on the ambitious three-year-long Dark Star project its goal to develop a new AI

    00:00:54 aided approach to solving these computer intensive physics problems leveraging the the laboratory's fastest and most powerful high- performance computer systems and advances in generative learning models the darkar inverse design framework helps researchers discover complex optimized science and engineering design solutions by working in a revolutionary new

    00:01:20 [Music] way we live in a three-dimensional world but now go to a higher Dimension four dimensions five six these are abstract spaces we really can't even imagine but mathematically they exist the curs of dimensionality basically says that as you add more parameters to your problem the harder it is to kind of explore that problem space imagine shining a a

    00:01:45 flashlight to try to find something it's extremely dark and you can only look in one place at a time as the number of Dimensions goes up it gets very hard to look and find things this is where machine learning comes in we're kind of at a cusp in terms of the curs of dimension a lot of the problems we've explored it seems that not all of the parameters

    00:02:04 actually matter in terms of some of our end goals in those cases you can almost rule some of those parameters out machine learning techniques allow us to very rapidly search those High dimensional spaces and I guess you could say remove the curse of dimensionality inverse design is a very powerful methodology that has recently emerged from the fields of artificial

    00:02:29 intelligence and machine learning inverse design is where you take a final state that you want to create and you work backwards to solve for what the initial state is that's going to give that to you we have to kind of run backwards in time to help figure out things like what kind of geometry we should use what kind of velocity we should use even we can even play with

    00:02:50 things like additive manufacturing all of those things are things we can control in this inverse design framework to help us achieve a specific outcome to do this you need to combine both machine machine learning and physics and Engineering codes together and it's really the fusion of those two methods that allows inverse design we're generally operating spaces

    00:03:11 that there's so many knobs that we can turn that it's very hard to understand when turning one knob down and another Knob up how each of those kind of connect together into your overall objective that you're trying to achieve so the inverse design framework is really important in this case because it's kind of like a way to automatically self-tune some of those knobs where a

    00:03:33 human might not be able to fit all those relationships in their head typically when you design a machine to do something uh they're very simple shapes they're shapes that a human can draw and represent maybe even with an equation very simply a sphere a cube right things like that we tend to build machines out of those regular shapes with inverse design it's really letting you explore

    00:03:58 what the role of additional comp lexity is so what if instead of a sphere it's something with all kinds of features on it right whether it's a combustion chamber or the wing of an aircraft or a drone there's no reason that the simple shapes are best however up until now complexity was something that had to be put to the side it was simply too many variables to either simulate or think

    00:04:22 about or Draw with AI we can Embrace that complexity and we're realizing that when you Embrace complexity very often you can find dramatically better answers whether it be for aerospace engineering inertial confinement Fusion or other applications that we've looked at that involve time dependent phenomena we have the computational resources to run thousands of these

    00:04:48 simulations in detail unfortunately analysts don't have the time to look through each one of those thousands of simulations there is a great benefit in building what we call surrogate models and these surrogate models are machine learning models that have been trained from the physics simulations themselves and so what we will do is run tens of thousands or hundreds of thousands of

    00:05:10 physics simulations that becomes a training set for the machine learning the machine learning algorithm learns from those and develops its own understanding of what those physical laws are so one of the really cool things about these AI models is they can actually be run on smaller resources like a laptop or a you know typical workstation computer they don't have to

    00:05:30 rerun all of those computations they actually are going behind the scenes to have already computed the result and they'll be able to see the answers kind of in real time what I did on the darkar team is I wore the AI hat where I was generally fitting huge generative models to thousands of thousands of hydrodynamic simulation results so one of the special things

    00:05:54 we've done in the dark star project is really Empower analysts these AI models allow analysts to actually look at simulation scenarios for things they didn't actually simulate effectively interpolating between simulations that they've already run what we're looking at right now is a simulation result of a high velocity impact between two copper plates and what we're seeing right now

    00:06:14 is that we're at the initial time where the plates are about to have impact but as we progress time further we can see that there's an instability that grows from the peration that was on one of the copper plates this is actually a live m demo where we're actually using a machine learning model in the loop to change what some of the initial parameters were um of this copper on

    00:06:39 copper impact so you can see as we change some of these parameter definitions the instability um lessens and changes shape but we can actually even go back in time and see the initial result that resulted in that instability the analyst of the future right is going to set up a particular scenario that they're interested in and they're going to run a

    00:07:04 simulation but what's different is the analyst in the future will also have the simulation results for all the different pertubations that they could have done AI is notorious for sometimes having hallucinations or uh areas where it's just uh simply not effective for that reason it was very important to us in darkar that we have experimental validation of the predictions we making

    00:07:30 with these tools and so throughout the course of the project we conducted over a dozen campaigns mostly involving high explosive experiments where we tested the predictions being made by the Darkstar AI tools as well as our physics simulation tools the team was amazed at how accurate uh these models and codes actually are and how effective they were in predicting the outcome of of these

    00:07:56 experiments there's a lot of industries that can really benefit from it um obviously anything in Fusion research is very interesting to us there's other applications you can imagine like for example like car crash scaros right you can imagine that you can change things like the size of the door or the height of an a pillar on a car so you can imagine this two people are looking at a

    00:08:16 computer and one person over the shoulder says what about if you change this and the guy the guy at the computer says well I don't know so he changes it and then instantly you'll be able to see the ramifications the results on screen AI is going to be a big driver for us to get to that future one day in addition to forecast models that predict A hurricane's path scientists might be

    00:08:39 able to reverse engineer the exact mix of air currents and rising sea temperatures that caused a hurricane to form in the first place or how to perfect a Fusion fuel capsule starting from the desired energy output and moving backwards in time to reveal optimal designs leading to the development of renewable energy fusion power plants AI assisted research and

    00:09:06 design can help us get there for more information on the darkar program visit Science and Technology review