WEBVTT

NOTE Replacement captions. A disputed theorem-count cue is excluded pending machine-generated reconciliation. Cues after the exclusion are retimed by -7.890 seconds.

1
00:00:00.000 --> 00:00:02.640
Water and air touch every human life.

2
00:00:02.940 --> 00:00:10.660
Yet water scarcity affects about two billion people, and outdoor air pollution is linked to millions of premature deaths each year.

3
00:00:10.960 --> 00:00:17.230
Both problems are materials-limited, and both are corrupted by the same predictable computational error.

4
00:00:17.530 --> 00:00:24.520
Universal machine-learning potentials learn tidy bulk crystals where every atom has many regular neighbors.

5
00:00:24.820 --> 00:00:34.710
But the membranes, sorbents, and catalysts that purify water and air function at pores, metal centers, and single-atom sites—environments with fewer neighbors.

6
00:00:35.010 --> 00:00:40.080
There, the energy surface is systematically softened by 15 to 60 percent.

7
00:00:40.380 --> 00:00:48.770
A softened barrier makes the wrong membrane look selective, the wrong sorbent look stable, or the wrong catalyst look fast.

8
00:00:49.070 --> 00:00:54.020
Across millions of candidates, tiny systematic errors flip rankings.

9
00:00:54.320 --> 00:00:58.530
Experiments chase false priorities while better materials stay on the bench.

10
00:00:58.830 --> 00:01:07.910
Corrected binding and barrier energies overturn raw rankings in desalination membranes, atmospheric-water sorbents, and lithium-selective frameworks.

11
00:01:08.210 --> 00:01:20.280
The same correction recovers activation energies for low-temperature exhaust catalysts, soot filters, and volatile-organic-compound oxidizers, separating real activity from artificial promise.

12
00:01:20.580 --> 00:01:28.080
The correction measures error as a field over local coordination, anchored to three observables and pinned to zero in bulk.

13
00:01:28.380 --> 00:01:35.030
Analytic forces are added at runtime, keeping molecular dynamics nearly as fast as the base model.

14
00:01:35.330 --> 00:01:44.500
Across 36 blind model-material pairs, corrected predictions correlate at 0.906 with zero fitted parameters.

15
00:01:44.800 --> 00:01:54.100
A membrane ranking is supported only when its local environments fall inside the measured domain, and a metastable sorbent phase is flagged honestly rather than sold as predicted.

16
00:01:54.400 --> 00:01:58.660
The failure mode is coordination-specific, not climate-specific.

17
00:01:58.960 --> 00:02:08.130
One measured correction-and-verification layer can raise discovery reliability across water, air, batteries, and direct air capture.

18
00:02:08.430 --> 00:02:11.500
The goal is predictions laboratories can trust.