- 135comments
- 185comments
- 33comments
- 221comments
- 58comments
- 11comments
- 45comments
- 19comments
- 242comments
- 71comments
- 237comments
- 35comments
- 9comments
- 17comments
- 51comments
- 168comments
- —discuss
- 147comments
- 24comments
- 13comments
- 16comments
- 3comments
- 89comments
- 174comments
- 72comments
- —discuss
- 148comments
- 15comments
- 280comments
- 137comments
Despite the power of topology optimization, it has very little adoption in industry. We believe that this is due to large barriers to entry, in the form large capital overhead from very expensive software licenses and the cost of HPC hardware. This is where Mechanical Mind comes in.
We plan to bundle our topology optimization software with compute services offered through AWS in a simple pay-as-you-go format. Users will be able to parameterize their design problems via a CAD plugin, then simply upload the compute job to our servers. Upon completion, the result is returned and the user is billed.
Typically, most topology optimization is based on a continuous approach, however we tackle this problem with a new discrete combinatoric method. This simplifies our solution search space and allows us to infer information about what sorts of solutions exist where. Also, as quantum computers mature, this solution approach directly maps to qubits, allowing near instantaneous solving by quantum methods.
Here's a simple proof of concept demo where a set of optimal cooling fins are designed to minimize the core temperature of a thermal body. Note that since this was run on a consumer laptop, the breadth of the search was limited for performance reasons. With more computational power, the quality of the solutions will improve.
https://imgur.com/a/T08nA
Any feedback or questions are appreciated! I'll be dropping in and out respond to them as they come.
How big is the market? And how easily users will switch given if they are comfortable with the existing software?
Excellent questions!
Engineeing services have 167 billion dollars of annual revenue in the US, and 773 billion globally. Because this software augments engineers' productivity while simultaneously improving the quality of their work, we expect to be able to achieve an appreciable percentage of that revenue number at full maturity.
Current adoption of topology optimization software in industry is negligible for the reasons mentioned earlier. Only the largest players in the market have access to it, and even then it's only applied to a limited domain of problems. Our goal is to bring this technology into wider adoption in industry by making it more financially accessible. We are aiming at small and medium sized engineering firms.
We also plan to integrate this software with the CAD platforms which our users are already familiar with, by offering a selection of plugins for these existing software packages.
First, I really like your startup, best of luck with it :)
In that case, why it's applied only to a limited domain ? It's not a financial issue, the fixed costs are already paid ? Or in most cases capacity if full and it would require expensive purchases ?
Thanks!
Most licensing is done on a per-seat basis for existing solutions. Additonally, the firms capable of swallowing the initial capital cost tend to be quite large and segmented into multiple divisions, e.g. Boeing or General Electric. Different divisions have different priorities within the company, different operating budgets, etc. Often a division's revenue won't be enough to justify the investment.
Pardon my ignorance of Computational topology design. You parameterize the problem and define a "cost" or "reward" as well somehow in the CAD plugin?
The short answer is yes.
Simulations via the finite difference method or finite volume method (depending on the problem type) are used to evaluate the fitness of each candidate solution on the server.
Using the plugin, the user can select from a number of objective functions, e.g. minimize mechanical stress in this area, maximize heat conduction in that area, via the native selection features of their CAD program. The plugin then performs meshing of the geometry, applies boundary conditions (e.g. forces, thermal loads), bundles it all together with information about the objective functions, and then ships it off to the server for evaluation and optimization. The user can also select from number of solution validity constraints to ensure compatibility with the manufacturing process that they have in mind.
Hello, is it something like Tosca by Dassault? Thank you and good luck.
Thank you!
It's analagous in that both are topology optimization solutions, however we are targeting a different market segment and also offering this as a networked service.
Hello again, another question, if you like. Are you trying / targeting industry-specific critical components already?
We're currently focusing on a broader application at the moment, since our primary goal is to bring this technology to the wider market. Our initial focus is on optimization of heat transfer and structural mechanics problems, however.
Why not just general analysis-as-a-service? Tosca is already a somewhat niche product within analysis software.
FE and such analysis programs cost $X000/seat, but selling relatively few seats. How will you make money selling 'per simulation' or hourly compared to licenses?
Definately an interesting concept
Thanks for the feedback!
Engineering analysis is already a pretty much solved problem. Topology optimization, however, has a lot of benefits that have yet to be realized, and it's much more well suited to the SaaS model because of how computationally intensive it is, since the analysis needs to be performed many thousands of times.
Because we outsource the capital cost of server hardware to AWS, we can scale linearly with demand which allows us to accept lower margins than our competitors. Our strategy is to make up for lower margins by reaching a much larger audience in industry than our competitors.
Why won't the big guys don't attack that market today ? and why won't they attack you and win when you start show hints as a credible threat ? Is that because you use discrete search, hence have lower compute costs ?
Topology optimization is still a new and emerging technology and it's viewed somewhat as an exotic engineering technique. Most software vendors recognize this and figure it makes better business sense to charge exorbitant prices to major players while focusing on their core products elsewhere. Also these vendors have been in the business of selling software directly, rather than as a service, for nearly half a century now and have yet to enter into the software as a service market. It remains to be seen how willing or able these companies will be to pivot to a software as a service model, especially given the size and age of their codebases. (Ansys was designed for punchcards!)
Also, yes, we do go about solving things pretty differently from existing players. Other solutions use very old, very large finite element solvers which were never really intended for this sort of thing in the first place. Our discrete finite difference approach allows us to make a lot of performance optimizations and iterate very rapidly through the search space. For example, we don't need to remesh each iteration and we can also look more directly at the nature of the search space, letting us make better decisions about where to traverse.
How will users trust your results, particularly in safety critical applications?
What are the main limitations of your approach?
How do you ensure that the parts you generate can actually be fabricated?
How do you plan to find your customers?
Good questions!
We use standard and well proven simulation techniques that have existed since the first half of the 20th century. We are however also verifying mesh convergence on each compute job, as well as verifying the output of the solver.
Topology optimization is inherently very computationally intensive by nature, though thanks to advances in computing that's becoming less of an issue.
A big component of topology optimization is the application of solution constraints. Users simply choose their manufacturing process, and then the software will mark all solutions that can't be manufactured by that process as invalid. So for example, when optimizing an injection molded part with a 2-piece mold, the optimizer will throw out any solutions which feature overhangs or internal cavities.
We're planning on riding the current buzz surrounding AI and automation in general. We plan to launch a social media campaign highlighting the effectiveness of topology optimization by comparing the performance of parts designed by our software versus those designed by a human competitor. Once we gain our initial userbase, growth should be straightforward once our business partners prove the advantage of using our software in industry.
How big is your team and what are their backgrounds?
Do you have a specific industry/customer you'll target first? Are they already using engineering services firms for this work, and can you do it cheaper/faster?
Are they really on social media? I would have thought that these people would be more likely found on a trade-show floor. This sounds like a pretty sales-heavy market, how will you get past the relationships a lot of incumbents will have built up over time?
Or is your plan to go after new entrants / give access to firms who want it but can't get it yet?
How do you prove your part is better than the human designed one, what's the key metric for 'better' for your customers ?
=======================================================
For the record, I think this is a really cool idea and probably the future of a lot of the downstream end of mechanical/process engineering, but execution is going to be everything.
I'm working with a lot of optimization right now in computer vision applications, and we're seeing the compute restrictions come down pretty quickly around this stuff too. It would be nice to have a general optimization engine somewhere I could just dump my problem into though.
Good questions and thanks for the feedback!
We're currently focused on the development of the software and getting a shippable product.
Our target demographic is small and medium sized engineering firms, who then go on to use our software assist in the design of products for their customers. The social media campaign is simply a fast, cheap means of establishing an initial presence despite being currently unknown. It will also allow us to reach smaller, more agile players quite well. But yes, our goal is simply to allow widespread adoption of this technology rather than uproot the current solutions being used in limited amounts by very large firms.
Essentially, we aim to do for topology optimization what Microsoft did for operating systems.
"Better" depends on the problem at hand, though typical measures of performance are cheaper, lighter, stronger, etc.
You're right that this is certainly a computationally expensive field, however our value proposition to potential customers makes up for that. Our business partners will have a clear market advantage, being able to produce higher performing designs while having greater engineer productivity than their competitors.
Suppose that for whatever reasons, automating mechanical design doesn't stick with your customers. What would you do then?
Good question.
Although that would quite frankly be pretty surprising, given the immense value we offer our customers, the next step would be to spin off the technology we've developed to an existing topology optimization software vendor. We're breaking new ground with this software, and some of our developments would undoubtedly be highly valuable to established companies in this area.
Cool idea and even cooler business model. In terms of design and product planning, I'm curious about how many prospective users you've talked to and what pain points they identified. Have you demo'd your prototype for any of them?
Absolutely! One of the other issues that hasn't been mentioned already but seems to come up a lot is the relative lack of different solution constraints available in these existing packages, preventing the use of optimization beyond parts intended for a limited set manufacturing processes. It is extremely difficult to create solution constraints from a software development perspective when working with a continuous solution approach, and yet it is trivial to do so with a discrete approach. This gives us a major advantage and allows us to generalize our software for more engineering problems, letting us reach wider adoption.
Most of what we have right now would still be considered in the "bench test" phase, but we're making progress towards something which could be demonstrated in industry. Our primary focus at the moment is making it easier to import and export CAD geometry to and from the optimizer.
Do you have evidence that your target customers are willing to go for a pay-as-you-go model? My personal experience with companies in the design and engineering space is that they hate non-fixed costs for tools and are generally willing to shell out larger sums of money up-front to "own" a tool (or contracted tier of use) outright rather than worrying about how much it will cost them over time as they create more designs/simulations.
Good question!
"Pay-as-you-go" is definitely open to interpretation, and whether we offer billing on a per-job, monthly, yearly, or contractual basis is still something that we're discussing. In the end, we're probably going to go with a mixture of these options, though right now our primary focus is on building a good product.
Most engineering software is already licensed on an annual basis, so recurring billing is definitely quite standard within engineering software tools.