Hello,
I’m working on a project called TRECH:
https://github.com/Geckos-Ink/trech
It is still very alpha, so I don’t want to sell it like a finished scientific tool. It is more a skeleton that I’m trying to make real step by step.
The idea is to build an environment around Geant4 where I can define experiments, run them, collect data, and then use statistical / torch-based training to infer behavior on larger and larger scales.
So the direction is to avoid more as possible:
“write a formula for everything and simulate only the result”.
It is more:
“start from essential elementary particle / material / interaction data, collect enough statistics, then try to train or infer what happens at higher scales with the smallest amount of pre-written laws possible.”
Of course this does not mean “ignore science math knowledge”. Some things should stay as hard constraints. For example, currently I’m not trying to make a neural network rediscover basic relativistic particle behavior or fundamental conservation rules. Those are still too fundamental parts to being predicted.
Formulated physical and chemical laws are essential in any case for tests validation.
The interesting part for TRECH is the rest: when a system becomes too big to describe manually with clean formulas, can I still build a pipeline where the lower-scale simulation produces data, and then torch/statistical models help infer the behavior at higher scales?
TRECH: Because I don’t think AI will replace humans’ work, I think it can replace Universe’s work.
So Geant4 remains the base where Geant4 makes sense. Then TRECH tries to build layers above it:
-
scripted experiments;
-
reproducible runs;
-
JSON outputs;
-
provenance of what was actually executed;
-
statistical aggregation;
-
torch-based inference models;
-
future scale-to-scale prediction;
-
support to PubChem APIs
-
and later a visual environment to create, inspect and execute experiments.
For the moment the project is heavily AI-agent developed, because I currently cannot type (and, honestly, think) enough. So yes, there are surely blind spots, imprecisions, wrong assumptions, maybe parts that are too optimistic or just badly written (but in the opposite, some issues smoothly solved that I would find blocking).
I’m working anyway because I want to first give the project a general structure. Then the physics boundaries, validation, documentation and architecture can be corrected with more clarity.
At the moment the two current milestone names are Sputnik and Apollo.
Sputnik is the first launch. The first small thing that goes up and proves that the base works. For TRECH this means the minimal core: Geant4-backed execution, scripted experiments, reproducible data, first outputs, first checks, first structure.
Apollo is the effective goal. It is not only “more examples”. Apollo is the future lab environment: a UI with 3D viewing where experiments can be set up visually, configured, executed, inspected, compared, and then used as data sources for analysis or training.
I share this with Geant4 community for an initial (shocking, I presume) confrontation, especially:
-
where the Geant4 boundary should be;
-
what should never be moved into a statistical/AI layer;
-
what kind of validation would make sense first;
-
whether a scripted + torch/statistical + future 3D UI environment around Geant4 is a reasonable direction;
-
and which existing Geant4 projects I should study before I build too much in the wrong way.
-
In a more generic physics and chemistry laws point of view, where test scenarios are tricked by too much pre-assumed formulas and constants.
(Again) TRECH is alpha, incomplete and probably full of things to fix, and with a lack of right external libraries to import.
In case of git pull requests, currently I may be a little slow to handle them, I need also a serious collaborative-ready structure.
Thanks for reading this far. If you just skipped right to the end, you’re a cheater.