Nonlinear constraints to BoFire (integration with acqf + tests + tutorials) by MrzvskK · Pull Request #740 · experimental-design/bofire

MrzvskK · 2026-03-04T17:02:42Z

Motivation

This is an early attempt, I am sure there will be a few changes needed along the way. But to start with something, so Nonlinear constraints (NonlinearInequalityConstraint and NonlinearEqualityConstraint) were defined in BoFire's domain model but were not integrated into the optimization pipeline. The constraints were extracted from the domain but never passed to BoTorch's optimize_acqf, causing the optimizer to ignore them and propose infeasible candidates.

This PR:

Fixes the constraint integration pipeline in acqf_optimization.py to properly pass constraints to BoTorch
Corrects a critical bug where the is_equality flag was inverted (all constraints were treated as equality constraints)
Enhances torch_tools.py to properly format constraints for BoTorch compatibility
Adds feasibility verification in botorch.py to ensure proposed candidates satisfy all constraints
Provides tests and a tutorial demonstrating usage

Have you read the Contributing Guidelines on pull requests?

Yes

Have you updated `CHANGELOG.md`?

I will do that later

New Test Suite

Created tests/bofire/strategies/test_nonlinear_constraints.py with 16 comprehensive test cases:

Constraint Types Tested:

Single and multiple inequality constraints (2-5 simultaneous)
Equality constraints
Mixed equality/inequality constraints

Optimization Scenarios:

Different acquisition functions (qLogEI, qLogNEI, qUCB)
Tight constraints (small feasible regions)
Boundary conditions (initial data on constraint manifold)
Single-variable constraints
Empty feasible regions (proper error handling)

Results:

All 16 nonlinear constraint tests passing
352/354 total strategy tests passing (99.4% pass rate)
All integration tests passing

Tutorial

Created docs/tutorials/basic_examples/nonlinear_constraints_basic_usage.py demonstrating:

Competing reactions optimization with nonlinear constraints
Constraint definition using string expressions
Initial feasible sampling
Multi-objective optimization with constraints
Constraint verification utilities

Verification Steps

To verify the changes work:

# Run nonlinear constraint tests
pytest tests/bofire/strategies/test_nonlinear_constraints.py -v
 
# Run full strategy test suite
pytest tests/bofire/strategies/ -v
 
# Run tutorial example - they also exist in .qmd - after revision we can delete .py files 
python docs/tutorials/advanced_examples/nonlinear_advanced.py
python docs/tutorials/advanced_examples/nonlinear_constraints_maximizing_yield.py

Add files via upload

…_Reaction_Optimization.qmd Co-authored-by: Johannes P. Dürholt <johannespeter.duerholt@evonik.com>

- Added filter_candidates_by_constraints() to filter feasible points - Added get_constraint_violations() to calculate violation amounts - Exported new functions in api.py - Added comprehensive tests

…torial runs smoothly

LukasHebing · 2026-03-05T10:14:28Z

Hi @MrzvskK,
nonlinear constraints are not handled by the standard BoFire optimizer, but the GA handles those:
https://experimental-design.github.io/bofire/docs/tutorials/advanced_examples/genetic_algorithm.html

Does this help?

MrzvskK · 2026-03-05T10:43:10Z

@LukasHebing, this PR adds functionality to the BotorchOptimizer to handle nonlinear constraints.

LukasHebing · 2026-03-05T11:25:46Z

@LukasHebing, this PR adds functionality to the BotorchOptimizer to handle nonlinear constraints.

Ok, great :)
I just wanted to make sure that this is not overlooked

MrzvskK · 2026-03-17T19:12:02Z

I fixed most of the issues, except of the last failing test, it has something to do with the conda.

MrzvskK · 2026-03-20T13:37:50Z

@jduerholt hi fixed all the bugs, should be ready to merge

MrzvskK · 2026-04-10T08:57:35Z

@jduerholt Hello, this is a quick reminder, I wanted to check if you approve the changes or not.

jduerholt · 2026-04-10T09:28:04Z

Ah sorry, did not see that the tests are now passing, I will review over the weekend or beginning of next week ;)

MrzvskK · 2026-04-17T08:22:53Z

@jduerholt kind reminder, did you have time to take a look at the PR?

jduerholt · 2026-04-17T12:25:18Z

@jduerholt kind reminder, did you have time to take a look at the PR?

Sorry, too much travelling, I will do it over the weekend or the latest on Monday.

jduerholt

Hi @MrzvskK,

thank you very much for the PR. I went over the actual implementation (ignoring the tests) for now and provided inline feedback. Regarding the botorch based suggestions, I also created an issue there: meta-pytorch/botorch#3280

Best,

Johannes

jduerholt · 2026-04-20T11:46:52Z

        return hessian_expression

-    def __call__(self, experiments: pd.DataFrame) -> pd.Series:
+    def __call__(


From my perspective, this is highly botorch dependable code, for this reason, I propose to move it to utils/torch_tools.py, where also the other constraints are prepared to be botorch ready. Then we also do not need the option here to pass torch ensors into the function etc, and it is much cleaner to read. You can have a look there how the ProductConstraint is build up.

jduerholt · 2026-04-20T11:49:17Z


    type: Literal["NonlinearEqualityConstraint"] = "NonlinearEqualityConstraint"

+    def is_fulfilled(self, experiments: pd.DataFrame, tol: float = 1e-6) -> pd.Series:


Honestly, I would not support them directly out of the box in bofire, as they are not naturally supported in botorch. If somebody wants to do something like this, he or she should do this breakdown into individual nonlinear constraints by hand.

jduerholt · 2026-04-20T11:50:29Z

            constraints.NonlinearEqualityConstraint,
        ]:
-            return False
+            return True  # was False


I would only support the NonlinearInequalityConstraint and would the rest be handled by the user.

jduerholt · 2026-04-20T11:51:40Z

            if self.local_search_config is not None:
                if has_local_search_region(domain) is False:
-                    warnings.warn(
+                    logger.info(


Why this change from warning to logger, the logging mechanism in BoFire is somehow not well implemented, so I would like to keep it as warning for now. We need an overhaul there ...

sorry, I forgot to revert it.

will do that now.

jduerholt · 2026-04-20T12:30:52Z

+                    )
+                    # Use object.__setattr__ to bypass Pydantic's frozen model behavior
+                    object.__setattr__(self, "batch_limit", 1)
+                if self.n_restarts != 1:


I do not think, that we need n_restarts ==1, it should also work with more restarts, but it will run them sequentially. So this can go, I think.

jduerholt · 2026-04-20T13:31:22Z

            return OptimizerEnum.OPTIMIZE_ACQF_MIXED
        return OptimizerEnum.OPTIMIZE_ACQF_MIXED_ALTERNATING

+    def _get_nonlinear_constraint_setup(


Hmm, I do not understand why this is necessary, currently we use our random strategy as initical candidates generator, and the random strategy can already handle nonlinear constraints via rejection sampling. Why not keep it as is?

The motivation for _get_nonlinear_constraint_setup is that once we pass nonlinear_inequality_constraints to BoTorch, it validates feasibility in tensor space with its own convention/tolerances, and ICs produced via BoFire rejection sampling can still be rejected there. If we decide not to pass nonlinear constraints into BoTorch and rely purely on BoFire’s rejection sampling, then I agree we can drop most of this setup and keep the implementation simpler. I will double check the RandomStrategy, since it might be doing the same thing.

I will explore get_initial_conditions_generator

jduerholt · 2026-04-20T13:34:10Z

            domain, constraint=LinearEqualityConstraint
        )
-        if len(nonlinear_constraints := get_nonlinear_constraints(domain)) == 0:
+


I also do not understand this, this should all be handled in the original implementation using the random strategy as generator for the feasible candidates.

jduerholt · 2026-04-20T13:51:57Z

    def has_sufficient_experiments(
        self,
    ) -> bool:
+        """Check if sufficient feasible experiments are available.


Hmm, I am not convinced that this is a good idea, from my perspective there is no need that past/historical experiments that were just used to train the surrogate models should fulfill the constraints. Often we have usecases, where new constraints are added over the course of the project, and this should keep working.

What was your intention for adding?

I can revert this change.

jduerholt · 2026-04-20T14:00:40Z

-        If the domain is compatible with polytope sampling, it uses the polytope sampling to generate
-        candidate samples. Otherwise, it performs rejection sampling by repeatedly generating candidate
-        samples until the desired number of valid samples is obtained.
+    def _ask(self, candidate_count: int) -> pd.DataFrame:


Hmm, I do not get why this change was necessary, can you elaborate a bit?

jduerholt · 2026-04-20T14:02:26Z

+            # The constraint should evaluate to <= 0
+            def make_constraint_callable(c):
+                def constraint_fn(x: Tensor) -> Tensor:
+                    # c.__call__ expects x with shape (batch_size, n_features)


I would move all the logic on handling torch tensors of different here.

MrzvskK and others added 20 commits January 20, 2026 22:05

Add files via upload

612e272

Merge pull request #1 from MrzvskK/MrzvskK-enting-tutorial

82760b5

Add files via upload

trailing-whitespace

276e30b

Merge branch 'experimental-design:main' into main

a57d154

Update docs/tutorials/advanced_examples/EntingStrategy_Suzuki-Miyaura…

7824dd1

…_Reaction_Optimization.qmd Co-authored-by: Johannes P. Dürholt <johannespeter.duerholt@evonik.com>

Merge branch 'experimental-design:main' into main

01dd371

Add helper functions for nonlinear constraints

62e92b1

- Added filter_candidates_by_constraints() to filter feasible points - Added get_constraint_violations() to calculate violation amounts - Exported new functions in api.py - Added comprehensive tests

updated nonlinear.py and reverting api to the earlier version

b5b1092

delete unnecesary test

e28f1fd

update

ff9ad1d

added smoke test and changes yml file

92ee319

wokring nonlinear constraints

a10ff2d

nonlinear constraints, tests and an example

302cae2

edits to the nonlinear constraints, making sure all tests pass and tu…

343e760

…torial runs smoothly

small edits to nonlinear package

61a27a5

tutorials to qmd format

a7cb501

tutorials to qmd

669157f

tutorials to qmd

0c0404d

Merge branch 'upstream-main' into nonlinearC

8220c8b

not create conflicts deleting enting tutorial

67fbaa2

jduerholt self-requested a review March 5, 2026 14:04

MrzvskK added 6 commits March 16, 2026 21:42

addressed some errors

164dc81

should fix quarto error with torch generator

d390ef1

changing nonlinear.py to work for pd.dataframe

31bdd72

added torch generator

4235d09

trying to fix the remaining tests

c9ff32b

another fix

a116486

MrzvskK added 3 commits March 18, 2026 11:10

need to restart the tests

333f7c4

small edit

ad017b0

some fixes to the failing tests

5a5925a

jduerholt mentioned this pull request Apr 20, 2026

[FEATURE REQUEST]: Extend project_to_feasibile_space_via_slsqp to nonlinear constraints meta-pytorch/botorch#3280

Open

1 task

jduerholt reviewed Apr 20, 2026

View reviewed changes


		type: Literal["NonlinearEqualityConstraint"] = "NonlinearEqualityConstraint"

		def is_fulfilled(self, experiments: pd.DataFrame, tol: float = 1e-6) -> pd.Series:

Conversation

MrzvskK commented Mar 4, 2026

Motivation

Have you read the Contributing Guidelines on pull requests?

Have you updated CHANGELOG.md?

New Test Suite

Tutorial

Verification Steps

Uh oh!

LukasHebing commented Mar 5, 2026

Uh oh!

MrzvskK commented Mar 5, 2026

Uh oh!

LukasHebing commented Mar 5, 2026

Uh oh!

MrzvskK commented Mar 17, 2026

Uh oh!

MrzvskK commented Mar 20, 2026

Uh oh!

MrzvskK commented Apr 10, 2026

Uh oh!

jduerholt commented Apr 10, 2026

Uh oh!

MrzvskK commented Apr 17, 2026

Uh oh!

jduerholt commented Apr 17, 2026

Uh oh!

jduerholt left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Have you updated `CHANGELOG.md`?