GRRRR part 2

On a sunny day (Wed, 01 Jun 2022 15:38:20 -0700) it happened John Larkin <jlarkin@highland_atwork_technology.com> wrote in snipped-for-privacy@4ax.com:

There was a headline in scidaily.com today about if AI designed stuff should be patented and by whom, or something

I know I probably draw fire if I say that all that spice is a dead end road at least for many things.

Neural networks, done some programming with that, we are like that.. could be the future in electronic design too. But a NN (Neural Net) is trained by building and testing things. Even if you spice on a supper computah there is NO guarantee the circuit will even work in reality.

I play with Raspberry Pis these days.. Even those are very much unobtainable due to chip or other shortages.. Would be nice to have LT spice running on ARM processors (maybe there is a port already?) There must be a break even point between AI an Spice type simulations for electronic design? Wonder how far that is away. For the rest I leave the problem solving to my neural net, have not touched spice in years? It is nice for filters that need a lot of repeated math scribbling, but there are other filter design programs. Many things do not have a good spice model...

It's sure not dead for people who design real electronics. What I want is for Spice to run on an Nvidia parallel compute engine, 200x or so faster than on an Intel cpu.

Are NNs anything but an academic toy?

Sometimes things work exactly as Spiced. Good engineers can usually expect when the sims aren't to be trusted.

There have been attempts to use computers to actually design circuits, or at least to optimize values in a given topology. They tended to be ludicrous failures.

It's strange that our brains, evolved to be hunter-gatherers, can design electronics.

That's the main hazard, not having dependable part models. That's a serious problem when using RF-type parts large-signal time domain.

We finally finished our big laser modulator chassis. The amplifier/fiducial board got to rev C, and we avoided D by adding a MiniCircuits SMA DC block in one of the cables. You can't sim this fast stuff; just guess and etch.

I guess one can plug an ssd into a PCIe slot. I need new PCs and it would be cool to have a fast D: drive on them, mostly for Dropbox.

They're great for flagging possibilities for later evaluation. IIRC they're used in drug discovery for that reason. For control systems, not so much.

It's not at all strange that our minds, being made in the image of the Maker, can do that. ;)

There's a lot of that going around.

Cheers

Phil Hobbs

(Who is in the middle of a board spin for our nanowatt photoreceiver, partly for that reason.)

Cross-probing large .raw files even stored on an SSD is still slow cuz they use a compressed format, to make it snappy you have to first convert the data file:

which is also slow. Or with your new machine you can buy a bunch of RAM say 32 or 64 gig and make half of it a RAM drive dedicated to sim data which also speeds up probing considerably.

Worked for me. Didn't make spice run any faster, but it did make mousing round in the stored waveforms much faster.

On a sunny day (Thu, 02 Jun 2022 06:33:32 -0700) it happened snipped-for-privacy@highlandsniptechnology.com wrote in snipped-for-privacy@4ax.com:

Many new medicines have been designed by AI now, it is worth keeping up to date on science by reading sciencedaily.com But also technical things, like airplane wings, what not. How did I learn? From looking at circuits, building and trying those. Big advantage for me is that I had to do fast fault finding in very complex systems and also simpler ones, so have seen thousands of designs, and HAD to grasp how those worked to be able to fix those. So you get an idea of the latest state of the art and what works and what not (when it fails) This is exactly how AI is trained, say for object recognition for military - or diagnostics for medical applications.

Looks neat, what's in that metal box on the right?

Yes, I have some eval boards, also lots of small boards from China for cheap. About triacs.. ever used an ACS108S? Third time one blew up in my Whirlpool washing machine Does not like spikes on the mains it seems, ordered 10 for 4 $ from ebay.. Very strange if you look at the datasheet it wants a negative drive, so takes plus as ground sort of thing

"This makes cross probing large circuits with huge simulation data files interactive."

It's the term the authors of the LTSpice wiki used themselves in the link you didn't read.

It means "You click on the node in the schematic you want to see the simulation plot for, and then the data you want to see appears in the window that shows the plot, or vice versa by clicking alt + left on a label in the plot pane and it will highlight the associated node"

• NN anything more than an academic toy? Ummm, they are integral computational entities in many products that require object detection, and pattern recognition. Many autonomous vehicles use them. I have worked on various airborne threat detection systems that have ANN classifiers.

I developed ANN (Artificial Neural Net) based object detection and signal processing back in the early to mid 1990s. One application that I can talk about is object detection and classification in airline luggage. The quality of the ANN was and still is directly related to the number of training sets you throw at it. Coming up with sufficient number of training sets was a challenge. Fast forward 10 years and the explosion of the web - images by the gazillions. They are more mainstream than you can imagine. J

I used an optimizer for some chip designs, it was very very good for things like choosing the size of the transistors in flipflops for best toggle frequency per supply current, and optimising a low pass filter for best noise and in-band error-vector-magnitude and stop-band rejection etc. all at the same time. It did way better than I could have. The trick was to write a script that runs the right simulations and results in an expression (or several) that correctly describes how well a circuit meets the goals. Once you've done that, it can twiddle the knobs much better than any human, and I don't mean because it could do it faster and spam the simulations across a thousand CPUs whilst you could look at only one at a time, it was also better in that it could remember many sets of parameters that were good in various ways, and combine them more efficiently than a human. It was a company internal tool and they will surely have kept it that way.

Sounds like a super useful tool. I did something similar for optimizing plasmonic nanoantennas 15 or so years ago, and like yours, it found good solutions that weren't at all obvious. So I'm a fan of the general approach.

Of course, that sort of thing has been done automatically since the

Numerical optimization based on merit / penalty functions has been well known for 200 years, since Gauss iirc. That's a far cry from actual computer-based design.

Even lenses, which you'd think would be a natural application, have been resistant to fully-automated design--it's all about finding a suitable starting point.

There are various approaches that modify topologies, of which the best known are genetic algorithms.

Cheers

Phil Hobbs

Yes. The part about " it could remember many sets of parameters that were good in various ways, and combine them more efficiently than a human" sounds very much like a genetic programming algorithm, which are very good at improving from a valid starting point.

Joe Gwinn

Not as far as I know. The point of genetic algos is to change the topology, not just the values. Unless I'm misunderstanding, Chris's optimizer was the usual sort that tweaks parameters to minimize some penalty function. Most of those remember previous values too--for instance my usual go-to algo, the Nelder-Mead downhill simplex method ('amoeba()' in Numerical Recipes). For N variables, it keeps N+1 sets.

I like Nelder-Mead because most of the things I need to optimize are either discontinuous themselves, like the number and placement of rectangular boxes of metal in a nanoantenna, or else need to be constrained to physically realizable values, as in a filter design code where the component values need to be positive. (I usually use mirroring to constrain that sort of thing, which avoids the tendency of the simplex to collapse along the discontinuity like water along a curb.)

Cheers

Phil Hobbs

It sounds like multi-dimensional slope-descent to me.

These are very good at finding local optima for a given design. Design from scratch requires finding global optima, something that slope-descent isn't very good at. Simulated annealing and genetic programming might have more luck.

CH

I actually used non-linear multi-parameter curve fitting in my Ph.D. work back around 1968. It did rely on a finite continuous data to create the surface that it crawled across. I used the Fletcher-Powell algorithm rather than Marquardt, but there are plenty of others.

That does sound right.

You can do the same thing with a list of components to generate a simple figure of merit for an application. Can include such esoteric issues as cost (book, real estate, process).

Don't stock market analysis programs try to do this in real time? Muddying their own pool . . . . .

RL

Sure. My EM simulator can optimize on literally anything expressible in its input files.

The issue isn't figuring out how good a design is, it's generating good ones from a blank sheet of paper.

Cheers

Phil Hobbs

Simulated annealing or simplex are pretty good for that sort of thing. Then conjugate gradients once you get somewhere near an optimum.

You may never be sure you have the global optimum but it will probably find a solution that is better than any human can in a reasonable time.

Optical lens designs have been pretty much solved now at least in the domains that I frequent. There could only be a handful of truly weird configurations remaining that haven't been tried already.

The last interesting one in terms of being very different to orthodoxy was Willstrop's three mirror telescope (no chromatic or sphereical aberration, precise focus and very fast).

I expect the odd novelty still lurks in the shadows. The search for eyepieces with ever more wide angle views remains the Holy grail - they are getting a bit ridiculous now with some offering 120 degree AFOV.

Modifying the topology is best done by humans. Optimising the components against some library of available materials and shapes is now the domain of sophisticated ray tracing programs. Zemax is probably the best known: