Empirical, Evolutionary Chemistry
Copyright (c) 1997, Forrest Bishop,
All Rights Reserved
Interview of Forrest Bishop
by Bill Spence A verbatim transcription.
December 10, 1996
B: Interview with Forrest Bishop, nanotechnologist. What's
your idea about empirical evolutionary chemistry?
F: Chemical synthesis.
B: Tell me about it.
F: Today new chemicals are found by -- the main methods in
[solution-based] chemistry involve a multistep procedure that takes place
over a period of days that involves distillations and ... a typical experimental
chemist will spend anywhere from hours to days trying a new synthetic
pathway, and then running the compound, distilling the compound...
B: To do what?
F: To discover a new compound.
B: OK, we're talking about discovering a new compound.
F: To characterize its properties.
B: OK, we're talking about trying to find new materials here?
F: Yeah, new chemical compounds, this is a generalized view
of how they're synthesized. They'll have some idea of what they're looking
for, and some idea of how to synthesize it. They want to take an existing
compound and stick a new group on it and see what that does to its properties.
So they will come up methods of cooking this compound in such a way that
this new group is added onto some existing compound to make a new type
of compound which then they have to isolate, distill and place in little
vials and glasses and things to do that on.
B: You're saying that ...time
F: Exactly. It takes a long time to do this.
B: OK, so what does evolutionary chemistry mean? What does
it do differently?
F: What we should be able to do, once we have the ability
to do positional control of the chemistry, with the nanomanipulators,
tiny robot arms, is create new compounds and test them at rates of millions
per second. The reason I say that is that, if you consider a robot arm
that moves at one meter per second, that picks something up and one second
later places it a meter away-- now imagine a robot arm that moves at one
meter per second, but it only moves 10 nanometers, zooming(?) ___ it picks
something up and moves it over 10 nanometers and reacts with something
over here. [So] at 10 nanometers at a meter per second will only take
it 10 nanoseconds to do that. So you've caused a chemical reaction of
your choice to occur in 10 nanoseconds, say, just as an example. Of course,
there's no reason to limit yourself to only one arm, one manipulator,
in this hypothetical nanotech chemistry lab.
B: So, how many arms are we talking about, just approximately?
F: Oh, let's make it a million, why not?
F: So you have a million arms--no, let's just say a thousand,
to keep it...
B: Doing different tasks?
F: Doing some the same, and some different.
B: You've got a thousand different experiments going on at
F: You have a thousand experiments going on at once.
B: And how fast will the results come out?
F: Each experiment is taking place at a rate of, let's say,
a thousand per second, to be conservative. In a thousandth of a second,
one of these arms should be able to assemble some new compound. Each of
these thousand arms is doing that, so you have a million new compounds
being formed in one second. Each of these compounds -- let's say you're
looking for some new behavior from this compound that you build--
B: Such as...? Superconducting? Room temperature?
F: Right. That would be a more complex structure. We'll go
further into that in a minute. We'll start with the simplest ideas, the
simplest compounds that are too complex to do quickly with current methods,
but interesting to do if you want to make [novel] compounds. Perhaps we're
looking for biological activity, or perhaps we're looking for catalytic
activity. Some specific property that we want this new compound to do.
So now we have a million compounds that we created in this one-second
interval, and we want to test each of those and find out which one of
them does the job the best. So that job--let's say it's a catalytic reaction
that we want this new compound to perform better than what we have already.
So we may have started with some existing compound that we already knew
did the reaction, or maybe not. Let's say we did; we started with some
compound that we already knew kind of worked, and we made a million variants
on that. We can then test each of those, each one of those "homemade"
new compounds ainst the reaction that we want it to perform by taking
the compound and moving it to another test chamber and allowing that reaction
to occur, or that process, whatever this new compound is supposed to do...
B: Looking for specific properties?
F: Yeah, and measuring its performance, and that then becomes
what is called the fitness parameter. You then take from all of those
million compounds that you tested, you pick the best ones and you feed
that information back to the software that is controlling the robot arms.
B: Tell me about the software.
F: Each arm has to have some control system to it, to position
the arm and move it and know what reactive group to grab, to move.
B: What kind of feedback software are we talking about? Where's
the smarts of it?
F: There's the control software for the arm, which is actually
what you're evolving--the control software for the arm and also the control
software for [the feedstock or reagent injectors]. Because if you feed
a different atom in, obviously, you're going to make a different compound.
B: Is this where the evolutionary chemistry comes in?
B: How's that work?
F: What you're actually evolving is software, but the software
is a description of the chemical compound, this chemical compound that
you're trying to make. So the nice thing about positional [-controlled]
chemistry is what you make depends up on what the software is that's controlling
the robot, so by changing that software, you're changing what you're making.
B: So each few nanoseconds, you change the software on every
robot arm, correct?
F: Depending on what the information was. Typically, in a
genetic algorithm, as they are applied today to software, you will actually
breed different sections of code together to create a new form of the
B: So the software is evolving.
F: Yeah, the software is evolving, exactly. The chemical,
the physical chemical is evolving, too. Its description is embodied in
the software. This is the difference between what we do now and what we
can do with nanotechnology, because the description of an object can be
coded into software.
B: The object...
F: Yes. You can consider a chemical compound to be a packet
of information in and of itself, but that packet of information can also
be expressed in a different way, in this case, in lines of code.
B: Evolving lines of code.
F: Evolving lines of code.
B: So, when you evolve lines of code, you're actually describing
evolving chemical compounds.
F: Yes. It's a back-and-forth procedure. Part of the evolution
is done in the software, part of it is done by physically containing the
chemical and seeing what it does. That process is what gives you the feedback
from the physical world back to the software and gives you the fitness
parameter, which is saying the measure of how well that particular compound
did the job you were looking for.
B: And this is the empirical part?
F: Yes. This is the experiment.
B: And this has, obviously, tremendous advantages over doing
molecular modelling, because you're cutting out lots of processes, like
trying to think of what the compound should be, to begin with, before
you model it?
F: Yeah, it's nice to have something as a precursor, to start
with. That cuts down your amount of search space that you need to look
through, but it's not strictly necessary. You could start with a water
molecule and just start adding things onto it and evolve it that way.
It would take longer, but it would still go somewhere.
B: Tell me about, if you have a specific goal in mind, and
you're doing a million experiments a second, looking for a particular
compound that would do something you specifically had in mind. Is it time
for the superconducting part?
F: OK, there's a new class of high-temperature superconductors
that are composed of various clays, actually, basically, that, by mixing
together certain elements--oxygen, yttrium, copper, boron--those kinds
of elements--and cooking them, they self-assemble into very intricate
structures in which layers of atoms with certain seed atoms placed in
very special, specific spots--all by self-assembly--there's no one human
involved in this, since it was discovered quite accidentally, it would
be nice to be able to repeat that with more direct control over the growth
process of what is essentially a crystal and be able to experiment with
different spacings and different types of atoms and things that maybe
wouldn't have self-assembled. So, if you want to make a superconductor
that has a very high radiation resistance, or a very high tensile strength,
or simply make one that's a single crystal, which the ones we have now
aren't--they have grain boundaries where you lose performance; it lowers
either the critical temrature or the critical magnetic field, that would
potentially be possible if it were a single crystal. So, the incentive
of being able to manufacture a crystal that's perfect would be, in and
of itself, very nice, but more importantly, to be able to evolve a better
one than could have been developed by simple self-assembly methods, one
would want to use this kind of empirical, evolving system. In this case,
the fitness test would be a little clamp, say, that you put this structure
in and test its superconductivity.
B: You could have though it was just constructed?
B: And test its superconductivity?
F: Yes, and that number would be fed back as the fitness
parameter for the software that built that superconductor, and if you
have a million of these, you can take the best ones, say the best thousand,
and cross-breed their code lines.
B: Whoa, and just do a real serious shotgun effect quickly,
as massive experiments in a very short period of time.
F: Massively parallel experiments.
B: And here's the real power of the idea. This is the real
power of having massive parallel experiments happening quickly.
F: Very quickly. And the interesting thing about genetic
algorithms is they veer off and lead to solutions that are not obvious
to a designer, things we never would have thought of.
F: Sometimes they're so wild that we don't even understand
how they work, some of these software programs.
B: Really? Do you have an example of that?
F: There was, I believe, in the very early days of GA's,
genetic algorithms, a fellow with a simple program to sort names alphabetically,
a nothing program, very small program. And then applied a genetic algorithm
B: This is evolution.
F: Yes, after it evolved inside the computer-- what the computer
does is take -- he basically fed in lines of code from existing seedstock
addressing program, name-arranging, and the computer, the program that
did the evolving would pick lines of code and assemble them into a new
little program that would arrange names alphabetically, and then the fitness
parameter was how many steps did it take it to do that, starting with
a random list of names and turn it into an ordered, alphabetically listed
list. So, the programs that would do that in the shortest amount of time
were judged to be the most fit for the process. But the interesting thing
was, when it finally levelled out at a certain size code, it was not only
a non-obvious solution, but the last I heard, they weren't even able to
figure out exactly how it worked.
B: Oh, so it evolved something that was much more efficient
and so compact, so novel and compact, it wasn't obvious how it actually
F: Yes. I'm sure they figured it out, but at the time, when
they first looked at it, it didn't seem to make sense, yet it did the
B: So, the idea of evolutionary chemistry uses powerful software
technique to have feedback systems and figure how to evolve novel chemical
F: That's right. Things that we never would have thought
B: And do it very quickly, very fast, on a huge, massive
F: And you notice there's no requirement in this procedure
for molecular modelling.
B: Right. So you don't have to think about---
F: In a way, it's doing the thinking for you.
B: Right. Instead of spending a week trying to figure out
one molecule, and then molecular model it to see if it goes together,
this thing just says, "Forget that. We're just going to do a shotgun
F: Yes. It is a shotgun approach.
B: That keeps evolving. It's a smart shotgun approach, on
a massive scale, is that correct?
F: Yes, yes. It is quite analogous to how evolution of life
works, which is very similar, where you're taking snippets of various
programs that were the most fit and most successful, to consider those
programs as organisms, consider their code lines as snippets of DNA, and
you take snippets of DNA from two different organisms, software, and you
combine them--like you take half of the code from one program and half
from another-- and you combine them to make a new program...
B: Code can have sex.
F: That's right. Code has sex. That's right. And it doesn't
just have to have just two sexes; it can have three or four or a thousand.
B: Code is kinky?
F: Very kinky. It can have as much sex as you want.
B: And so you __________ it in the same kind of _____________
that evolution is used to kill
F: Yes. Genetic algorithms have been around for quite some
time, 15 years, I think, commercially, so it's a known procedure.
B: So you have a smart, evolving shotgun procedure to develop
[novel] compounds with particular parameters that you're looking for that
nobody's ever thought of before.
F: Yeah, nobody _could_ have thought of before.
B: Nobody could have thought of before.
F: That's right.
B: Sounds like a powerful technique. How big is this lab?
F: It depends upon what you're making. For a simple chemical
that's -- well, simple in our terms--say 100,000 Dalton chemical -- the
lab might be the size of a matchbook.
B: The whole lab doing a million experiments and testing
the experiments and evolving?
F: That's right, that's right. And that's just to keep a
conservative number on it.
B: That sounds like nanotechnology to me. The power of nanotechnology.
F: Yeah, that's just to keep a conservative number. It would
probably be smaller,
B: You were mentioning...
F: ...depending upon the types of tests you do. Some tests
require high magnetic fields, for instance, which do require large equipment
to generate. So it depends upon the actual fitness parameters that you're
B: So, instead of a matchbox, it might be a box of cards.
B: But a cigar box would cover just about anything you could
B: So, we're talking about an incredibly powerful laptop
F: That's right.
B: Interesting, interesting. You were mentioning earlier
some very unusual requirements for a device called the Starseed /Launcher
( for the launching rails.
F: Yeah, not the [rails], for the conductors of that have
very stringent requirements as far as superconductivity, radiation hardness
and tensile strength, and so this would be an example of what we were
talking earlier about, evolving a superconductor. This is a specific example.
B: Because the material you're talking about just does not
exist yet, right?
F: Yeah, there's materials that would do some sort of a job
of it, but there's so much room for improvement that why not go for the
best you can get.
B: OK, so you've got those three parameters, where you need
superconducting, you need high tensile strength, and you need radiation
hardness. Would that be a three-tiered testing device?
F: Yes, you could consider each of those as a fitness parameter,
so you --depending on what you want to make, you can vary the weight of
those parameters and -- for instance, if you want -- out of a hundred,
let's say you want 50 of it to be superconductivity and 25 of it to be
radiation hardness, and 25 to be tensile strength, that's when you ________.
If you vary those weights, and weight it more towards tensile strength,
that's going to give you a different class of compounds. So in other words,
you can specify what you want first, and then by putting that in as a
fitness parameter, you can evolve what you need.
B: You could just stack on as many parameters as you need,
depending on what kind of material you need for a particular job.
F: Yes. The software gets complicated quickly, but yes, you
B: And all the while, you're doing a million experiments
F: That's right.
B: So, in a day's time, you can test more novel materials
than have ever been tested in all of history, is that what you're basically
F: No, I'm saying you can do that in a minute.
B: Oh, so we're talking about a very powerful...
F: Make that a New York minute.
B: So, we're talking an extraordinarily powerful way of coming
up with new materials?
F: Yes, sir.
B: I like the sir part. Is there anything else on this particular
F: Well, then, we've talked about fairly small molecules
and then somewhat more complicated crystal structures for high-temperature
superconductors. But then, it need not stop there. There's already ongoing
work and, I believe, commercial application, of evolvable hardware in
the computer industry. Genetic algorithms are already being used to design
very large-scale integration of computer chips and to do the layout of
a printed circuit board. If you consider building a circuit board, you
have a lot of components and you have a lot of leads going between those
components. It actually becomes a very complex problem to figure out the
best way to run the leads between the different components, with the minimum
number of crossovers, vias through the board from one side to the other...
B: And distance.
F: And distance, and crosstalk between wires. There are several
fitness parameters on that.
B: So the premise for this whole concept is already up and
F: Oh, yeah, well-established; that's right.
B: ...and mature.
F: It's still developing, but, yes,
B: So people will be able to plug right in.
F: There are already commercially available packages that
do various types of genetic algorithm searches for various types of problems.
This one in particular I mentioned because it is similar to evolutionary
chemistry in that you're actually evolving a physical item via the software
that describes it.
B: Run that by me again. That's an important point.
F: Yeah, a printed circuit board is a physical item. There
are certain requirements, or fitness parameters, that you're shooting
for, the ones we spoke of--the minimum number of vias, the minimum number
of crossovers, the minimum number of layers, and so forth. And, of course,
minimize the size of the board. There are already packages that do this
in software. The software will describe the layout of a PC board. In this
case, the computer is simply talking to itself, because it is able to
fully characterize that PC board in software, but it is still -- the purpose
is to build a physical object and in principle, it could describe a board,
build it, and then test it that way, too. In fact, maybe the board is
complicated enough, if you start to talk about crosstalk and ringing and
all these other problems that come up in high-speed boards, maybe they
do actually build boards at some point and test [them], trying to minimize.
So, anyway, these packages will construct a description of the board and
then, say, meure all lengths of all conductors on that board. That can
be done on software easily.
B: So this is the type of software that could be put directly
into characterizing the materials that you build a million a second.
F: Well, it's an analogous approach, and you obviously can't
use the same software. But the point is, we already have a very good analogy
of this operating in the real world. It's just this kind of search.
B: Anything else you can think of? It think we've covered
F: No, the other reason I bring up evolvable hardware, which
is what the field is called already, is that you don't have to stop at
simple compounds and structures. Once we are building nanoscale devices
of some complexity, it would be very desirable to evolve those as well.
So, this same kind of procedure can be used... The reason I bring up PC
boards being evolved and integrated circuits being evolved. That field's
called evolveable hardware already. If we can evolve chemical compounds
and superconductors, why not evolve entire systems, nanotechnological
devices. If you have the computational power available, and you have the
ability to conduct either thousands or millions of experiments per second
on this hardware, why not evolve mechanisms and nanoscale electronic devices
just as we do microtechnological devices today?
B: Really? In other words, do a shotgun approach on nanoscale
B: In a matter of hours, you'd have billions of different
experiments, and the very best ones according to your parameters would
be popped out.
B: And no molecular modeling, no trying to preconceive of
what the best one's supposed to be, and see if it fits together?
F: You still want to do some preconception so that you can
see_____ bring in some kind of seed to seed it with.
B: Right. But the best machine for the job evolves and is
presented to you in a very quick fashion.
F: That's right.
B: And it's probably something that nobody would have ever
thought of, or could have thought of.
F: I would say, almost certainly, in 99.9% of the cases,
it would be inconceivable.
B: That's fascinating. That's almost -- you could consider
it some sort of artificial intelligence, in a sense, in a very narrow
F: It's directed evolution.
B: So powerful.
F: It's powerful, but it's -- boy, it depends on your definition
of intelligence, and there's so many definitions out there, that...
B: Right. Well, I guess one thing you could say is...It's
smarter than we are.
F: It's smarter than we are in the same sense a computer
is, though. A computer can multiply two numbers together a lot faster
than we can. And this can evolve hardware a lot faster than we can. So,
it's not conscious; it's not a conscious process; the conscious process
is our input along the way. We can, if we see that it's going off on a
tangent, what's called a local minimum. If it lands in some local minimum
that's sub-optimal, that you really don't want it to go to, you can bump
it out of that by changing a few lines of code or something so that the
thing wanders off and hopefully finds a better minimum. So it's a directed
form of evolution. There is conscious input into it, but the steps in
between are quite automatic.
B: So, to review, this is a way of shotgunning intelligently
millions of research, or researching millions of compounds for specific
properties that you're looking for and then identifying them as they present
themselves to you, and you can keep the process going on as long as you
wish and just unattended, as long as you wish, and who knows what may
come up through the mud?
F: That's correct.
B: And this is literally billions, or maybe trillions, of
times faster than doing it by hand in the lab.
B: Powerful new technique for coming up with not only novel
chemical compounds with specific properties that you want...
F: Materials and devices.
B: But materials and whole devices, wow.
F: Sure. The computer that's running it--you could, should
be able to evolve entire supercomputers. It's the same principle, and
it's a lot like what's already being done now in the semiconductor industry.
B: What if you were to use this thing to evolve a better
supercomputer that immediately went on-line to evolve a better supercomputer
that immediately went on-line to evolve -- that's...
F: That's something that will be done.
B: That's possible, isn't it?
F: It's already being done now. That's why computers are
getting so much faster than they were. Computers are designing the next
B: With this sort of smart ...?
F: I'm not sure how much of that that they use; I'm not quite
clear on how much of it is genetic algorithm, and how much of it is some
other type of technique.
B: In this particular scenario...
F: In this particular scenario, they will do it this way,
and it will be done this way.
B: In ten minutes, the world could get very scary.
F: That's right. When you're building _______ devices like
this, there's thermal limits on how much you could put together before
the heat load gets too great. You can only assemble so many atoms _________________
before the endothermic [heat of] reaction overtakes your cooling stuff,
so 10 or 12 minutes.
B: OK. I think we're wrapped. Done.
Copyright ©1967-2004, Forrest Bishop, All Rights Reserved