Thanks for this Marcus, > One could imagine coupling a physical simulation to a search procedure for > functional behaviors like memories and doorways. The detection > combinatorics would be challenging, assuming the physical simulation were > possible at sufficient fidelity, but perhaps could be constrained by virtue > of spatial locality.
Yes. I can’t bring to mind anybody who seems to be doing important work in this entirely within simulation, but Leroy Cronin in Glasgow is trying to combine highly parallel robot-maintained reaction vessels with pattern-matching computation and feedback, so see if he can search for properties and then extract chemical mixtures that will instantiate them. http://www.chem.gla.ac.uk/cronin/ Lee is a handful, and his group is the size of a small town (larger than some small towns in NM, I suspect), so one gets the sense of massive seiving for most-anything, with the hope that some fraction of that will remain of interest for longer than the time Lee is promoting it. The project is really different, though, from anything I have seen before. It begs to be integrated with modern AI, which has become quite flexible in what you are allowed to call salient, so one can search in very open-ended ways. > I don't know much about coarse-graining organic chemistry simulators. For > comparison, with molecular dynamics a billion atoms is possible (on a budget > of a few megawatts), but not for more than tens of nanoseconds. I've found > game physics engines like Bullet Physics are nice for coarse-grained models > because they are fast (optimized to graphics processors) and easy to > interleave control or detection logic. However, they couldn't (without more > work) decompose the space across memory domains of a cluster. Interesting. I don’t know much about molecular-dynamics simulations, which is deeply an expert’s game, though I guess most major universities have somebody in chemistry or biochem who specializes in it. If one is willing to go one level out, and ask which questions are hard at the level of network synthesis and search, taking reaction primitives as input data, the graph-grammar methods are becoming pretty sophisticated. The best I know of is the current state of the project that started with Peter Stadler but is now dispersed across German-speaking Europe and Scandinavia: http://cheminf.imada.sdu.dk/mod/ Much of what makes this a hard and interesting computational project doesn’t show if one merely wants to do chemistry. It comes up because they are trying to create a consistent representational system. This creates difficulties like deciding when two things are the same molecule; when two molecules arrived at through different pathways are actually isomorphisms of the same label set, etc. In random network-extension algorithms, this entails solving the graph-isomorphism problem a very very large number of times, and the underlying representational system must be provably well-defined. It is in coming up with representations that are more well-defined than SMILES or INCHI, and implementing most-modern isomorphism searches, that these guys are the furthest along. There is also a playground linked from their main page, though I am told they recognize their documentation may be a bit off-putting to people not used to wading into new systems. The current state of the graph-grammar project is several-fold: 1. Bond topology is present, and has been for some time. 2. Simple stereochemistry of carbon is now implemented, and less-simple stereochemistry that requires non-local propagation through a molecule to determine equivalence of representations is next to come. 3. Stereochemistry of metals, which will be the gateway to crucial mechanisms of metal catalysis, is planned. 4. There is a project, privately held, to port the entire Beilstein database of reaction mechanisms to graph-grammar representations, after which engines like this become an incredible tool. RIght now the bottleneck is usually manual coding of the mechanisms of interest. 5. There has been some discussion of raising pattern-matching above the level of atoms or local clusters, to inductively-defined patterns like crystal faces, but no serious attempt to formulate that problem yet. Even with the limited state of what they can do, they have achieved some tolerable comparisons against messy chemical systems, like the formaldehyde-addition network known as the “formose network”, and the HCN polymerization and hydrolysis system. Both are famously complicated, and both have long-standing interest to Origin of Life people, though their exact situation relative to planetary chemistry is easy to argue about. All best, Eric ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove
