WEBVTT

NOTE Narration transcript. Text is the script synthesized by minimax (speech-2.8-hd, voice English_expressive_narrator); each cue was synthesized as its own audio file and is timed by that file's measured duration. Verified at 115.7% of expected length, 125.3 wpm (median cue 131.3 wpm). Script source: public/videos/a-field-not-a-neural-net.vtt at commit 4641d96 (pre-overwrite narration prose)

1
00:00:00.000 --> 00:00:05.150
Okay, so Google DeepMind predicted 2.2 million crystals.

2
00:00:05.450 --> 00:00:06.380
Huge.

3
00:00:06.680 --> 00:00:13.920
But by late 2023, only 736 had been independently synthesized.

4
00:00:14.220 --> 00:00:17.710
The usual response is: build a bigger neural net.

5
00:00:18.010 --> 00:00:20.440
What if that is solving the wrong problem?

6
00:00:20.740 --> 00:00:27.040
Universal machine-learning potentials are already fast and pretty accurate for tidy bulk crystals.

7
00:00:27.340 --> 00:00:33.860
But real materials work at vacancies, surfaces, and transition states—places where atoms have fewer neighbors.

8
00:00:34.160 --> 00:00:40.130
There, errors jump, and ion-migration barriers can be underestimated by more than 60 percent.

9
00:00:40.430 --> 00:00:44.480
Here is the twist: that wrongness is not random.

10
00:00:44.780 --> 00:00:45.760
It has a shape.

11
00:00:46.060 --> 00:00:54.120
The correction measures an environment error field based on coordination number—basically, how many neighboring atoms each atom has.

12
00:00:54.420 --> 00:00:56.250
No second neural net.

13
00:00:56.550 --> 00:00:58.640
No per-system retraining.

14
00:00:58.940 --> 00:01:04.860
Three standard measurements anchor the curve: two surface energies and one vacancy energy.

15
00:01:05.160 --> 00:01:08.530
Bulk coordination is fixed at zero error.

16
00:01:08.830 --> 00:01:13.520
Then the field predicts a fourth surface—the one-ten facet—blind.

17
00:01:13.820 --> 00:01:21.160
Across 36 model-and-material combinations, predicted and measured errors correlate at zero point nine oh six.

18
00:01:21.460 --> 00:01:29.050
Flip that field, add it beside the existing model, and under-coordinated structures move back toward reference energies.

19
00:01:29.350 --> 00:01:34.120
The forces are analytic, so simulations stay physically consistent.

20
00:01:34.420 --> 00:01:43.420
In tests, nickel surface error dropped from 9.7 to 1.5 percent, while bulk lattice constants stayed unchanged.

21
00:01:43.720 --> 00:01:49.130
And this layer does something most AI pipelines cannot: it can say no.

22
00:01:49.430 --> 00:01:58.640
Lean four proofs check every quantitative claim and flag ranking inversions, already-converged cases, or physics outside the field’s domain.

23
00:01:58.940 --> 00:02:04.830
One proof even caught a rounding tie and corrected “27 improvements” to 26.

24
00:02:05.130 --> 00:02:15.230
That matters for battery ion hops, carbon-capture frameworks, ammonia catalysts, and perovskite stability—all controlled by under-coordinated defects.

25
00:02:15.530 --> 00:02:18.770
So the goal is not more confident predictions.

26
00:02:19.070 --> 00:02:25.370
It is fast predictions with measured corrections, certified boundaries, and a provable reason to stop.