Physics-informed ConvLSTM mass-balance loss (Pilar 2 x Pilar 3) — temporal_piml_loss • edaphos

3.10.0

Reference
- Capstone — six pillars integrated
- Cross-pillar campaign
- Pilar 1 — Causal AI
- Causal backdoor on br_cerrado
- Causal on 1 095 real WoSIS profiles
- Causal-discovery trio (expert × LLM × data-driven)
- LLM-KG benchmark (Gemma 4 × GPT × Claude)
- Annotation workflow (pre-annotation + Shiny review)
- Pilar 2 — Physics-Informed ML
- Parametric ODE + Neural ODE
- Pilar 3 — 4D Pedometry
- ConvLSTM on synthetic cube
- Real MODIS + NASA POWER + EnKF
- Pilar 4 — Foundation Models
- SimCLR / MoCo embeddings
- Pilar 5 — Active Learning
- Hybrid policy + physics gate
- SoilGrids-BR live
- Pilar 6 — Quantum ML
- ZZFeatureMap + Q-KRR
- Cross-pillar bridges
- Pilar 1 × Pilar 4 — Foundation IVs
- Pilar 4 × Pilar 6 — Quantum-foundation kernel
- Unified infrastructure
- Unified uncertainty API
- Case — end-to-end Cerrado benchmark
- Releases
- v2.0.0 — Quantum kernel × Foundation (Pilar 4 × Pilar 6)
- v1.9.x — Foundation embeddings as causal IVs + Sargan/Cinelli-Hazlett
- v1.8.x — LLM benchmark infrastructure + Shiny annotation tool
- v1.7.x — Capstone vignette + native-query calibration + causal discovery
- Changelog

Factory that returns a function(y_pred, y_true, driver) closure suitable as the physics_loss_fn argument of temporal_convlstm_fit(). The closure penalises SOC-dynamics predictions that violate the local-rate kinetics inferred from the Pilar 2 ODE fit.

Usage

temporal_piml_loss(ode_fit, weight = 0.1)

Arguments

ode_fit: An edaphos_piml_profile or edaphos_piml_bayes fit.
weight: Numeric; loss weight relative to the MSE term. Default 0.1.

Value

A function loss_fn(y_pred, y_true, driver) returning a scalar tensor/number (same shape agnostic to torch / R).