[UPDATE] The readme file got an update

README.md

@@ -103,17 +103,6 @@ Each trajectory was generated via a four-stage pipeline:
 
 ---
 
-## Key Findings
-
-Evaluating SchNet, PAINN, GemNet, DimeNet++, EquiFormerV2, MACE, NequIP, eSCN, and foundation models (MACE-MP-0, UMA-small, all fine-tuned) reveals:
-
-- **OOD errors are 1–2 orders of magnitude higher** than ID errors across all tasks and architectures
-- **ID accuracy does not predict OOD accuracy** — models that rank highest on in-distribution test sets frequently rank lowest on OOD test sets, and vice versa
-- **Force direction generalises well; force magnitude does not** — cosine similarities remain above 0.99 on OOD molecules, but magnitude and energy errors blow up, indicating the bottleneck lies in how local representations are aggregated into extensive energetic predictions
-- **Fragment Combination is the easiest** OOD task; Fragment Composition and Duplication are the hardest
-
----
-
 ## Data Format
 
 Data is stored in **ASE trajectory format** (`.traj`), readable natively with the Atomic Simulation Environment. Each frame contains: