Widen numpy constraint to <2.4

[pmrv]

Pull Request Template

Description of Changes

Relaxes the numpy pin in setup.py from numpy<=1.26.4 to numpy<2.4. Versions up to 2.3.x are empirically known to work; 2.4 is excluded because it breaks pyace, with a full diagnosis below.

Diagnosis: why numpy ≥2.4 breaks the energy output

PyACECalculator.calculate() builds its ASE results dict like this (src/pyace/asecalc.py:167):

'energy': np.float64(self.energy.reshape(-1, )),

self.energy is a 0-d array wrapping the scalar total energy from the C++ evaluator, so .reshape(-1,) turns it into a shape-(1,) array, which is then passed through the np.float64 constructor (not a cast of a Python float).

The behavior of that constructor changed in numpy 2.4:

• numpy ≤2.3: np.float64(array_of_size_1) collapses the array to a scalar np.float64. Since numpy 1.25 this has emitted DeprecationWarning: Conversion of an array with ndim > 0 to a scalar is deprecated, and will error in future.
• numpy 2.4: that deprecation expired. Scalar constructors now behave like astype for ndim > 0 inputs and return an ndarray (numpy 2.4 release notes, "NumPy scalar constructor behaviour for arrays and 0-d inputs").

So on numpy 2.4, results['energy'] becomes array([E_total]) instead of a float, and atoms.get_potential_energy() returns a 1-element array of the total energy rather than a scalar — the observed symptom. Verified empirically:

numpy 2.3.5: np.float64(np.array(-123.456).reshape(-1,)) -> np.float64(-123.456)   (scalar, ndim 0, + DeprecationWarning)
numpy 2.4.6: np.float64(np.array(-123.456).reshape(-1,)) -> array([-123.456])      (ndarray, ndim 1)

The same pattern affects free_energy on the same line pair, and energy / free_energy_std / energy_std in PyACEEnsembleCalculator (src/pyace/asecalc.py:308-315).

Second, independent numpy 2.4 breakage

pyace.radial.integrate() (src/pyace/radial.py:20) calls np.trapz, which numpy 2.0 kept only as a deprecated alias of np.trapezoid and numpy 2.4 removed outright — so radial-function integration raises AttributeError on 2.4 regardless of the energy issue.

Forward path (not in this PR)

Both are one-line fixes if support beyond 2.4 is wanted later: replace np.float64(self.energy.reshape(-1,)) with float(self.energy) (works on all versions and silences the existing DeprecationWarning), and np.trapz with np.trapezoid (available since numpy 2.0). Keeping this PR to the empirically validated <2.4 bound.

Checklist

• Code is well-documented.
• All tests have been run and passed. (dependency-pin change only; the C++ extension is not built in this environment — behavior difference verified against standalone numpy 2.3.5 vs 2.4.6)
• Relevant documentation has been updated if necessary. (no docs reference the numpy pin)

License Agreement

By submitting this pull request, I agree that:

• The code submitted in this pull request will be distributed under the Academic Software License (free for academic non-commercial use, not free for commercial use), see LICENSE.md for more details.
• The copyright for the code, including the submitted code, remains with Ruhr University Bochum. Ruhr University Bochum retains the right to transfer or modify the copyright.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_015Ui6Drmop9gZXo1fgn8yVg

---

Generated by Claude Code

[pmrv]

lgtm.merge