Roadmap and status of the six pillars of edaphos

Documentation-only landing page that summarises the scope of every pillar, its current implementation status, and the governing mathematical object. Use this together with the package overview — edaphos-package — as the starting point for navigating the API.

Details

Pillar	Namespace	Status	Governing object
1. Causal AI	`causal_*`	implemented	Structural causal model \(G = (V, E)\) with backdoor-adjusted estimand \(\beta_{x \to y}^{\text{do}}\) (Pearl, 2009); LM and BART estimators; real-data Cerrado application (`causal_cerrado_real_dag()`, v1.4.0) identifying direct effects on 1095 WoSIS topsoil profiles (MAP, tree cover, clay); LLM-driven Knowledge-Graph pipeline (`causal_kg_new()`, `causal_llm_extract()`, `causal_augment_dag()`) supporting Ollama / OpenAI / Anthropic; multi-extractor voting (`causal_llm_vote()`, `causal_llm_ingest_abstract_voted()`) with majority / weighted / intersection rules for consensus across backends; paginated corpus clients for SciELO / OpenAlex + `causal_corpus_deduplicate()`; resumable disk-cached ingestion (`causal_llm_ingest_corpus()` with `cache_dir` + `max_retries`); ontology alignment against a curated Cerrado vocabulary and live FAO AGROVOC SPARQL, including concurrent batched alignment (`causal_ontology_agrovoc_align_batch()`); bottom-up structure learning from horizon data via `causal_structure_learn()` (hc / tabu / pc-stable / mmhc backed by `bnlearn`) with bootstrap edge confidence; paper-scale persistence and audit via `causal_kg_save()` / `causal_kg_load()` (portable RDS edge-list), `causal_kg_to_turtle()` (W3C RDF 1.1 Turtle export with reified provenance), `causal_kg_rank_edges()` (multi-source ranking) and a `summary()` method.
2. Physics-Informed ML	`piml_*`	implemented	Pedogenetic ODE \(dy/dz = -\lambda_0 e^{-\mu z}(y - y_\infty)\) and Neural ODE \(dy/dz = f_\theta(z, y, \mathbf{x})\); Bayesian posterior over \((\lambda_0, \mu, y_\infty, y_0)\) via Laplace approximation and adaptive random-walk Metropolis (`piml_profile_fit_bayesian()`); Neural-ODE deep ensemble (`piml_neural_ode_fit_ensemble()`) approximating the predictive posterior through K independent trajectories.
3. 4D Pedometry	`temporal_*`	implemented	Stacked ConvLSTM (Shi et al., 2015) with seq-to-seq training, multi-step rollout and a mass-balance physics loss.
4. Foundation Models	`foundation_*`	implemented	SimCLR scaffold (`foundation_simclr_pretrain()`), MoCo v2 (`foundation_moco_pretrain()`) with momentum encoder + dictionary queue + raster-specific augmentations, and a planetary-scale tile pipeline (`foundation_tile_source_soilgrids()`, `foundation_tile_align()`, `foundation_tile_dataset()`, `foundation_moco_pretrain_tiles()` with Apple MPS / NVIDIA CUDA device dispatch, `foundation_moco_embed_raster()`); downstream fine-tuning API (`foundation_fit_classifier()`, `foundation_fit_regressor()`) with linear-probe and full fine-tuning regimes + a two-group learning-rate schedule; public pretrained weights distribution via Zenodo (`foundation_weights_list()`, `foundation_weights_download()`, `foundation_weights_load()`) with SHA-256 verification and an on-disk cache under `tools::R_user_dir("edaphos")`. The first published encoder, `edaphos-cerrado-moco-v1` (DOI 10.5281/zenodo.19701276, CC-BY-4.0), was pretrained on 50 000 16x16 Cerrado tiles (SoilGrids + WorldClim + SRTM). Honestly benchmarked in the v1.3.0 `case-cerrado-end-to-end` vignette against a ranger + gstat kriging baseline on 1212 real WoSIS profiles.
5. Autonomous Active Learning	`al_*`	implemented	Hybrid policy \(\pi(\mathbf{x}) = \alpha\,\tilde u(\mathbf{x}) + (1-\alpha)\,\tilde d(\mathbf{x})\) with PIML-backed gate; BatchBALD information-theoretic batch acquisition (`al_query_batchbald()`) with greedy log-det submodular optimisation over QRF tree-level predictions.
6. Quantum ML	`quantum_*`	implemented	Pure-R ZZFeatureMap (`quantum_feature_map()`) + quantum-kernel Gram matrix (`quantum_kernel()`) + kernel ridge regression (`quantum_krr_fit()`); Qiskit-backed VQE bridge (`quantum_hamiltonian()`, `quantum_hamiltonian_h2()`, `quantum_hamiltonian_organo_mineral()`, `quantum_vqe_fit()`, `quantum_vqe_exact()`) with three back ends — exact statevector, shot-based Aer with SPSA, and full IBM Quantum Runtime dispatch (`quantum_ibmq_submit()`, `quantum_ibmq_available()`, `quantum_ibmq_backends()`, `quantum_ibmq_least_busy()`) with M3 readout mitigation and ZNE gate-folding error mitigation; qiskit-nature bridge (`quantum_hamiltonian_from_pyscf()`, `quantum_hamiltonian_organo_mineral_nature()`, `quantum_nature_total_energy()`) that lifts a user-supplied XYZ geometry through PySCF RHF + FreezeCore + ActiveSpace + ParityMapper into a qubit Hamiltonian, with curated organo-mineral presets (formic-acid carboxylate, methanediol ortho-diol, Fe(III)–formate mineral binding).