Condition on material + particle physical properties, not learned embeddings #7

New Issue

2026-07-08T13:52:21+02:00

lbogner commented

2026-07-08 13:52:21 +02:00

(Migrated from gitlab.etp.kit.edu)

Condition the surrogate on material and particle physical properties rather than learned embeddings of material name / PDG code.

Material: Z, A, ρ, X₀ (radiation length), λ (interaction length).
Particle: mass, charge (mass is already needed to reconstruct secondary momentum).

Why: physical-property conditioning is the path to generalising across materials/particles instead of memorising the training menu (currently 2 materials, ~192 PDG codes). It also gives the MoE router meaningful continuous conditioning axes, and continues the "more physics into the net" theme. Related hypothesis (Jan, 2026-06-25): the peaked photon edep distribution may be a material-property effect that simplifies once the relevant properties are fed in. Worth comparing generalisation on a held-out material against the embedding baseline.

Source: Jan & Tobias meeting 2026-07-07 §6 (also the minicalosim-geometry open question).

Condition the surrogate on **material and particle *physical properties*** rather than learned embeddings of material name / PDG code. - **Material:** Z, A, ρ, X₀ (radiation length), λ (interaction length). - **Particle:** mass, charge (mass is already needed to reconstruct secondary momentum). **Why:** physical-property conditioning is the path to *generalising across materials/particles* instead of memorising the training menu (currently 2 materials, ~192 PDG codes). It also gives the MoE router meaningful continuous conditioning axes, and continues the "more physics into the net" theme. Related hypothesis (Jan, 2026-06-25): the peaked photon `edep` distribution may be a material-property effect that simplifies once the relevant properties are fed in. Worth comparing generalisation on a held-out material against the embedding baseline. Source: Jan & Tobias meeting 2026-07-07 §6 (also the `minicalosim-geometry` open question).

lbogner commented

2026-07-17 12:35:00 +02:00

(Migrated from gitlab.etp.kit.edu)

created branch 7-condition-on-material-particle-physical-properties-not-learned-embeddings to address this issue

created branch [`7-condition-on-material-particle-physical-properties-not-learned-embeddings`](/lbogner/giant/-/compare/master...7-condition-on-material-particle-physical-properties-not-learned-embeddings) to address this issue

lbogner commented

2026-07-17 12:35:09 +02:00

(Migrated from gitlab.etp.kit.edu)

mentioned in merge request !5

lbogner commented

2026-07-22 10:39:00 +02:00

(Migrated from gitlab.etp.kit.edu)

mentioned in commit 4d190696fe

mentioned in commit 4d190696fecce5d71a6ed81d589629093b3462fd

lbogner (Migrated from gitlab.etp.kit.edu) closed this issue

2026-07-22 10:39:02 +02:00

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: lars/giant#7