Results#

This page summarizes the measured outcomes and principal findings. Blank fields should be completed only after the corresponding summary files have been validated.

Index#

2D Materials Project Cases#

2D Materials Project benchmark cases#

Model

Device

Floating-point precision

Attacks

Epsilon values

Iterative attack steps

Random seeds

Cases per material

MACE-MH-1

CPU

float32, float64

FGSM, I-FGSM, PGD

0.001–0.005, 0.01–0.05, 0.1–0.5, 1–5, 10

1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100

42, 43, 44, 45, 46

850

UMA-S-1p1

CPU

float32, float64

FGSM, I-FGSM, PGD

0.001–0.005, 0.01–0.05, 0.1–0.5, 1–5, 10

1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100

42, 43, 44, 45, 46

850

CHGNet

CPU

float32, float64

FGSM, I-FGSM, PGD

0.001–0.005, 0.01–0.05, 0.1–0.5, 1–5, 10

1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100

42, 43, 44, 45, 46

850

Total

CPU

6 model/precision combinations

3 attacks

21 epsilon values

11 step values

5 seeds

2,550 per material

The 2D Materials Project dataset contains 20 materials, giving a total of 51,000 benchmark cases.

LiCoHPF Database Cases#

LiCoHPF database benchmark cases#

Model

Device

Floating-point precision

Attacks

Epsilon values

Iterative attack steps

Random seeds

Cases per material

MACE-MH-1

CPU

float32, float64

FGSM, I-FGSM, PGD

0.001–0.005, 0.01–0.05, 0.1–0.5, 1–5, 10

1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100

42, 43, 44, 45, 46

850

UMA-S-1p1

CPU

float32, float64

FGSM, I-FGSM, PGD

0.001–0.005, 0.01–0.05, 0.1–0.5, 1–5, 10

1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100

42, 43, 44, 45, 46

850

CHGNet

CPU

float32, float64

FGSM, I-FGSM, PGD

0.001–0.005, 0.01–0.05, 0.1–0.5, 1–5, 10

1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100

42, 43, 44, 45, 46

850

MTP

CPU

float64

FGSM, I-FGSM, PGD

0.001–0.005, 0.01–0.05, 0.1–0.5, 1–5, 10

1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100

42, 43, 44, 45, 46

425

MACE Model

GPU

float32, float64

FGSM, I-FGSM, PGD

0.001–0.005, 0.01–0.05, 0.1–0.5, 1–5, 10

1, 2, 3, 4, 5, 10, 20, 30, 40, 50, 100

42, 43, 44, 45, 46

850

Total

CPU and GPU

9 model/precision combinations

3 attacks

21 epsilon values

11 step values

5 seeds

3,825 per material

The LiCoHPF dataset contains 20 materials, giving a total of 76,500 benchmark cases.

For each model, precision, material, and random seed, the benchmark contains:

  • 63 epsilon-sweep cases: 21 epsilon values multiplied by 3 attacks.

  • 22 step-sweep cases: 11 step values multiplied by I-FGSM and PGD.

  • 85 combined epsilon- and step-sweep cases.

Outcomes#

Outcome

Stage

Result

Delta force

Post-attack

MACE and MTP break off to 106 – 109 after 1 min lattice

Delta force

Post-attack + relaxation

MACE and MTP remain at 108`eV/Å after perturbation of ~1% min. lattice, while other models ``<` 1eV/Å

Displacement

Post-attack

Smooth exponential increase as epsilon sizes increase

Displacement

Post-attack + relaxation

Models begin to diverge, MACE and MTP experience a greater initial spike at ~1% min lattice

Relaxation steps

Initial relaxation

CHGNet takes most steps ~250, MACE and MTP take the least

Relaxation steps

Post-attack + relaxation

Fill in

Neighbour Jaccard distance

Post-attack

Fill in

Neighbour Jaccard distance

Post-attack + relaxation

Fill in

Maximum coordination change

Post-attack

Fill in

Maximum coordination change

Post-attack + relaxation

Fill in

RDF L1 distance

Post-attack

Fill in

RDF L1 distance

Post-attack + relaxation

Fill in

Space-group change

Post-attack

Fill in

Space-group change

Post-attack + relaxation

Fill in

Symmetry-operation retention

Post-attack

Fill in

Symmetry-operation retention

Post-attack + relaxation

Fill in

Unique-site change

Post-attack

Fill in

Unique-site change

Post-attack + relaxation

Fill in

Key takeaways#

Finding

Question

Result

Recovery threshold

At what normalized epsilon does recovery begin to fail?

1% of min lattice parameter

Largest force response

Which model/attack/material produces the largest delta force?

MACE and MTP pretrained by LiCOHPF dataset

Slowest convergence

Which model most frequently reaches the 600-step limit?

CHGNet

Most recoverable model

Which model has the smallest post-relaxation changes?

Fill in

Least recoverable model

Which model has the largest persistent changes?

Fill in

Attack comparison

How do FGSM, I-FGSM, and PGD differ under matched budgets?

Fill in

Chemistry dependence

Do LiCOHPF and 2D materials exhibit different failure patterns?

Fill in

Topology transition

When do connectivity or coordination changes become persistent?

Fill in

Precision dependence

Do float32 and float64 support the same conclusions?

Fill in

Seed dependence

Are conclusions stable across seeds 42 through 46?

No abnormalities which suggest results are stable and reliable.

Contour exploration baseline

Are adversarial directions more damaging than contour motion?

Slight increase but many magnitudes less than deliberate attacks, suggesting that atoms merely moving away from equilibrium do not lead to failure modes compared to adversarial perturbations that maximize model loss.

Supercell dependence

Do conclusions persist for expanded periodic systems?

Fill in