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 117.3% of expected length, 123.6 wpm (median cue 134.2 wpm). Script source: public/videos/beyond-carbon-the-error-geometry-of-environmental-materials.vtt at commit 4641d96 (pre-overwrite narration prose)

1
00:00:00.000 --> 00:00:12.590
What do desalination membranes, methane catalysts, safer refrigerants, lithium recovery, PFAS cleanup, clean-air catalysts, and low-carbon cement have in common?

2
00:00:12.890 --> 00:00:15.650
Weirdly, the same atomic blind spot.

3
00:00:15.950 --> 00:00:18.840
The materials bottleneck is not climate-specific.

4
00:00:19.140 --> 00:00:21.490
It is coordination-specific.

5
00:00:21.790 --> 00:00:27.460
Universal machine-learning potentials learn tidy bulk crystals, where atoms have lots of regular neighbors.

6
00:00:27.760 --> 00:00:35.990
But useful chemistry happens at pores, surfaces, vacancies, and transition states—where atoms have fewer neighbors.

7
00:00:36.290 --> 00:00:41.750
Right there, these models can soften the energy landscape by 15 to 60 percent.

8
00:00:42.050 --> 00:00:44.140
That is not harmless noise.

9
00:00:44.440 --> 00:00:52.630
A softened barrier can make the wrong membrane look selective, the wrong catalyst look fast, or an unmakeable cement phase look stable.

10
00:00:52.930 --> 00:00:58.900
Across millions of candidates, tiny systematic errors flip rankings—and experiments chase false winners.

11
00:00:59.200 --> 00:01:04.960
The stakes are huge: water scarcity affects roughly two billion people.

12
00:01:05.260 --> 00:01:07.810
Air pollution kills millions every year.

13
00:01:08.110 --> 00:01:11.600
Methane drives about 30 percent of current warming.

14
00:01:11.900 --> 00:01:18.170
Clean-energy mineral demand could rise four- to six-fold by 2040.

15
00:01:18.470 --> 00:01:24.010
And cement alone produces roughly 2.8 gigatonnes of carbon dioxide annually.

16
00:01:24.310 --> 00:01:28.750
The correction's response starts with the useful part: the error has a shape.

17
00:01:29.050 --> 00:01:32.630
Three measured observables anchor an environment error field.

18
00:01:32.930 --> 00:01:42.520
Bulk coordination is fixed at zero error, and a constrained curve predicts corrections for under-coordinated environments the field never directly saw.

19
00:01:42.820 --> 00:01:52.460
Add that field at runtime, and simulations follow corrected energies and analytic forces—still about one hundred thousand times faster than quantum chemistry.

20
00:01:52.760 --> 00:02:00.900
In blind tests across 36 model-material pairs, predicted corrections correlated at 0.906.

21
00:02:01.200 --> 00:02:08.700
Then 77 build-locked Lean 4 theorems mark what is supported, out of domain, or synthesis-dependent.

22
00:02:09.000 --> 00:02:10.710
So this is bigger than carbon.

23
00:02:11.010 --> 00:02:21.840
One correction-and-verification layer can improve discovery across water, air, methane, refrigerants, minerals, PFAS, and cement.

24
00:02:22.140 --> 00:02:23.200
Measure the error.

25
00:02:23.500 --> 00:02:24.810
Correct the physics.

26
00:02:25.110 --> 00:02:26.340
Prove the boundary.

27
00:02:26.640 --> 00:02:30.058
Then send experiments the candidates actually worth making.