Epic: 2020 GDP/MINLP roadmap revisited — status audit and open ideas (meeting_20200527)

[bernalde]

Epic: Revisiting the 2020 GDP/MINLP roadmap (IDAES meeting slides, 2020-05-27)

Source: meeting_20200527.pptx — IDAES Subtask 3.1: Conceptual Design of Advanced Energy Systems (D. Bernal & I. Grossmann, CMU, May 27, 2020).
Audit: August 16, 2026, against Pyomo/pyomo@ce205a5fe (main).

This epic audits every improvement idea in that deck against today's Pyomo (code + literature), records what already landed, and tracks what is still open and interesting. Every open item now has its own issue — defects and broadly-wanted features upstream on Pyomo/pyomo, research-agenda items here on this fork. This issue is the index. Slide references as (sN).

---

Part 1 — Landed since 2020 (context, no action)

2020 idea	Where it lives today
Logical modeling layer	BooleanVar/LogicalConstraint (2020), core.logical_to_linear, contrib.logical_to_disjunctive (2022); GDP→MIP transformations handle logic internally
Hull perspective options (s3–4)	gdp.hull with FurmanSawayaGrossmann (default) / LeeGrossmann / GrossmannLee ε-approximations (Furman et al., COAP 2020)
Exact hull via cones (s4)	Partial: exact_hull_quadratic option in gdp.hull (landed 2026-03; quadratic-only, rotated-SOC form, per Gusev & Bernal Neira, arXiv:2508.16093) — see A4
Multiple big-M (s17)	gdp.mbigm (2022; [Trespalacios & Grossmann 2015]) — see A5
Between big-M and hull (s17-adjacent)	gdp.partition_disjuncts / gdp.between_steps (Kronqvist, Misener & Tsay, CPAIOR 2021) — restricted case only, see A7
Nested disjunctions (s10)	bigm/hull/binary_multiplication nested rework (2024, e.g. Pyomo#3213)
MindtPy single-tree LP/NLP B&B (s16)	mindtpy/single_tree.py lazy callbacks for cplex_persistent/gurobi_persistent — see C1
MindtPy FP + level/regularized OA (s12)	mindtpy.fp; add_regularization ∈ {level_L1/L2/L∞, grad/hess/sqp Lagrangian via PyNumero} (Kronqvist et al., MathProg 2020; Bernal et al., JOGO 2022)
Community detection merge (s15)	contrib.community_detection (Louvain, generate_structured_model()) — analysis-only, see C5
New GDPopt algorithms	gdpopt.ric, gdpopt.enumerate, gdpopt.ldsda (Ovalle et al., C&ChE 2025) — see B7
GDP model library (s11)	gdplib — active development at SECQUOIA/gdplib, see D1
Pyomo.GDP ecosystem paper (s18, s23)	Chen et al., OptEng 2022

In flight right now: non-rigorous crossed-bound detection (Pyomo#3947 → PR Pyomo#4013), LBB time-limit results (PR Pyomo#4010), GLOA convex cut slope (PR Pyomo#4011), LBB relaxed-NLP node solver (PR Pyomo#4012).

---

Part 2 — Open ideas, each now tracked in its own issue

Filed upstream on Pyomo/pyomo (defects & broadly-wanted features)

Item	Upstream issue
B1 — disjunctive FBBT/OBBT options silently ignored (regression)	Pyomo#4014
B7a — LDSDA missing from meta-solver dispatch (+ vestigial code)	Pyomo#4015
B9a — relative_gap() ZeroDivisionError / misplaced epsilon	Pyomo#4016
B9b — docstring advertises unimplemented algorithms	Pyomo#4017
(audit find) — gdpopt.enumerate TypeError on unbounded/float-bounded integers	Pyomo#4018
A8 — solver-native indicator-constraint transformation	Pyomo#4019
C1 — MindtPy single tree on APPSI/contrib.solver	Pyomo#4020
B2 — LBB solver failures misclassified as node infeasibility	Pyomo#4023

A. GDP reformulations (pyomo.gdp)

• A1. First-class basic steps (s6) → A1: Register basic steps as a first-class gdp.basic_step transformation #3. Register gdp.basic_step + selection heuristics; wire into gdp.cuttingplane. Unimplemented in any OSS package.
• A2. Pseudo basic steps — Lagrangian dual bounds (s7) → A2: Pseudo basic steps — Lagrangian dual bounds for GDP #4. Papageorgiou & Trespalacios 2018; parallelizable GDP dual bounds; mpi-sppy analogy. Unimplemented anywhere.
• A3. gdp.mpec — GDP→complementarity (s5) → A3: gdp.mpec — GDP-to-complementarity reformulation and NLP heuristics #5. Forward direction absent everywhere; NLP heuristics/initialization angle (Schewe & Schmidt 2019).
• A4. Exact hull beyond quadratics (s4) → A4: Generalize exact hull beyond quadratics (cones, conic emission, reaggregation) #6. Exp/power cones (Bernal Neira & Grossmann, COAP 2024), kernel conic emission, reaggregated hull; lift mutable-Param freezing.
• A5. gdp.mbigm nested-GDP support → A5: gdp.mbigm nested-GDP support #7. Currently GDP_Error; sync with upstream maintainers on design before building.
• A6. gdp.cuttingplane modernization (s3) → A6: gdp.cuttingplane modernization #8. Nonlinear separation, shipped basic-step tightening, structural soundness, cut management.
• A7. General P-split formulations → A7: General P-split formulations #9. Kronqvist, Misener & Tsay, MathProg 2025 — general case implemented nowhere; automatic partition selection.
• A8. Indicator-constraint transformation (s3) → filed upstream: Support solver-native indicator constraints as a GDP transformation target Pyomo/pyomo#4019 (needs representation + interface support, not just a transformation).

B. GDPopt

• B1. Reinstate disjunctive FBBT/OBBT range reduction (s8) → filed upstream: GDPopt: calc_disjunctive_bounds and obbt_disjunctive_bounds are accepted but silently ignored Pyomo/pyomo#4014. Regression: options dead since the 2022 rewrite, GLOA cuts silently weakened, no-op test.
• B2. LBB rehabilitation (s9) → B2: LBB rehabilitation #10. Branching rules, non-XOR support, sane failure bounds, retire the 2022 "known issues" warning. (PRs Fix GDPopt LBB time-limit results Pyomo/pyomo#4010/Add LBB relaxed NLP node solver Pyomo/pyomo#4012 in flight; solver failures misclassified as node infeasibility filed upstream as GDPopt LBB treats solver RuntimeError as node infeasibility Pyomo/pyomo#4023.)
• B3. Logic-based branch-and-cut — "GDPopt-LBC" (s9) → B3: Logic-based branch-and-cut (GDPopt-LBC) #23. Node cuts + tree-wide cut pool; depends on B2/B1/B6.
• B4. Deeper Z3/SMT integration (s9) → B4: Deeper Z3/SMT integration in GDPopt #11. Complete term encoding, discrete presolve, unsat-core no-good strengthening, LD-SDA neighbor pruning.
• B5. Smarter set-covering initialization (s20–22) → B5: Smarter set-covering initialization for GDPopt #12. Max-coverage selection (deck's 8-process example: 2 vs 3 iterations); no newer literature variant exists — publishable.
• B6. Disjunctive cut strengthening (s17) → B6: Disjunctive cut strengthening and tightened-M cut placement #13. Kronqvist & Misener, OptEng 2021 — no OSS implementation anywhere; plus tightened-M cut placement (needs B1).
• B7. LD-SDA research extensions + gdpopt.ldbd → B7: LD-SDA research extensions and gdpopt.ldbd #14 (research); dispatch/vestigial bugs filed upstream: GDPopt meta-solver cannot dispatch to LDSDA (missing from _supported_algorithms) Pyomo/pyomo#4015. LD-BD per Liñán & Ricardez-Sandoval, AIChE J 2023.
• B8. Convexify-then-LOA ("Ruiz way") (s10) → B8: Convexify-then-LOA for nonconvex disjuncts (the "Ruiz way") #15. Per-disjunct convexification feeding LOA; middle ground to GLOA.
• B9. Meta-solver honesty → filed upstream: GDPopt relative_gap() raises ZeroDivisionError near zero primal bounds (misplaced epsilon guard) Pyomo/pyomo#4016 (relative_gap bug) and GDPopt docstring advertises unimplemented algorithms and omits implemented ones Pyomo/pyomo#4017 (phantom docstring features). A third audit find, the gdpopt.enumerate bounds crash, is gdpopt.enumerate raises bare TypeError for unbounded or float-bounded integer variables Pyomo/pyomo#4018.

C. MindtPy

• C1. Single tree on modern interfaces (s16) → filed upstream: MindtPy: support single-tree LP/NLP B&B with APPSI / contrib.solver persistent interfaces Pyomo/pyomo#4020 (APPSI callback exists; threads=1; solution-pool rewrite).
• C2. AD-based cut generation via PyNumero (s16) → C2: MindtPy AD-based cut generation via PyNumero #16. Numeric Jacobian rows → cut expressions; unify with the existing grey-box AD path.
• C3. Convexification/preprocessing layer (s16) → C3: Convexification/preprocessing layer for MindtPy and GDPopt #17. Coramin integration vs MC++ refresh vs in-tree FBBT/OBBT; home for B8 machinery.
• C4. Solver-interface QoL → C4: MindtPy solver-interface quality-of-life fixes #18. HiGHS QP routing, grey-box single tree, cyipopt duals, solution-pool surprise, fragile bound patching.
• C5. Community-detection decomposition — DeCODe-OA (s14–15) → C5: Community-detection-driven decomposition (DeCODe-OA) #19. Connect contrib.community_detection to OA/Benders; DECOA/DeCODe never landed as maintained software.

D. Ecosystem & papers

• D1. GDPlib refresh + archival paper + benchmark harness (s11) → D1: GDPlib refresh, archival citation paper, benchmark harness #20.
• D2. Archival MindtPy 2.0 software paper → D2: Archival MindtPy 2.0 software paper #21. Docs still cite PSE 2018; no consolidated citation for the FP/ROA/single-tree generation.
• D3. Documentation gaps → D3: Documentation gaps (GDP solving guidance) #22. LD-SDA vs enumerate guidance, disjunctive bounds, exact_hull_quadratic caveats.

---

Suggested first wave

Small & well-scoped (all upstream, PR-ready): Pyomo#4014 (revive disjunctive bounds), Pyomo#4015 (LDSDA dispatch), Pyomo#4016 (one-line gap fix), Pyomo#4017 (docstring), Pyomo#4018 (enumerate errors). Then fork-side #3 (A1) as the first transformation project.
Paper-sized projects: #4 (A2), #5 (A3), #23 (B3), #12 (B5), #13 (B6), #19 (C5) — each pairs an implementation with a publishable contribution; A2/A3/B6 are unimplemented in any OSS package per the August 2026 literature check.

Key references

• Kronqvist, Bernal Neira & Grossmann, 50 years of mixed-integer nonlinear and disjunctive programming, EJOR 2025 — current state-of-the-art survey.
• Chen et al., Pyomo.GDP: an ecosystem for logic based modeling and optimization development, OptEng 2022.
• Perez, Joshi & Grossmann, DisjunctiveProgramming.jl, JuliaCon Proc. 2023 — feature-parity checklist for A8.
• Full per-item citations in the linked issues.