Shulou(Shulou.com)12/24 Report--

Xin Zhiyuan reports

Edit: peaches

[guide to Xin Zhiyuan] Last week, a paper jointly published by A-Lab, DeepMind and other teams found that AI can create its own synthesis. Unexpectedly, a professor at the University of London found that there was a serious problem with the characterization of this material.

There is a serious loophole in the latest Nature paper of the DeepMind team.

Robert Palgrave, a chemistry professor from the University of London, revealed on the Internet that there were serious problems in the characterization of materials in the paper.

Even more outrageous, Palgrave found that AI had made three 90-year-old compounds and got the wrong ingredients.

The paper, "An autonomous laboratory for the accelerated synthesis of novel materials," published in Nature on November 29, is a collaboration between teams from UC Berkeley, Lawrence Berkeley National Laboratory and Google DeepMind.

In this paper, AI was realized in only 17 days, and 41 new materials were synthesized among 58 predicted materials.

Address: https://www.nature.com/ articles / s41586-023-06734murw

What is the problem with the characterization of materials? Let's find out.

Vulnerability Analysis Professor Palgrave analyzes his own views in the next thread.

Many "new" compounds are reported in this paper. The only feature they show is powder XRD, and there is no composition analysis.

But if they are good at XRD analysis, maybe it doesn't matter?

For clarity, if you are not used to viewing PXRD, the residual (red line) should be as flat as possible.

The residual is larger than most peaks.

Even if combined with other features, this cannot be a reliable improvement. But as the only form? Impossible.

The following picture is Mg3MnNi3O8. This is a "new" compound of hexagonal system, and its cationic sequence is very interesting.

The only problem here is that Mg6MnO8 and Ni6MnO8 are known compounds and are cubic.

In fact, the solid solution of these two compounds, probably written as Mg3MnNi3O8, was reported in 1995.

XRD enthusiasts please take a look at the PXRD pattern of "new" and "hexagonal" Mg3MnNi3O8 in front of the two tweets. It looks very three-dimensional, very similar to the Fm-3m space group.

We don't have raw XRD data, but it looks like it's probably a cube pattern, but it's actually a solid solution that was reported 28 years ago.

What about this MnAgO2? Fitting is also bad.

To make matters worse, it was reported in 2021 and has actually been structured by another high-throughput computing team and stored in the ICSD database (the database used in this article).

Let's take a look at how these two materials, different "new" materials.

They all look similar, with Sb, Pb, and O, obviously one with Hf and the other with Fe.

So, are there any known Sb, Pb and O compounds? In fact, Sb2Pb2O7 was mentioned in the report as early as 1933.

It has exactly the same pattern as the two "new" compounds mentioned above.

The collection code on ICSD is 24246 and Sn2Sb2Pb4O13.

Apparently, they successfully made three 90-year-old compounds without realizing and getting the wrong ingredients.

Another Sb2Pb2O7 is misrecognized.

The other is actually Sb2Pb2O7, which has the same pattern.

The author's latest response to the problems pointed out by Professor Robert Palgrave, UC Berkeley Professor Gerbrand Ceder responded:

Recently, our team published an article about A-Lab, an autonomous laboratory for the synthesis of target compounds driven by artificial intelligence. The purpose of this paper is to prove that autonomous agents can make decisions on how to synthesize a given material based on the history of text mining and ab initio thermochemical data (such as MaterialsProject).

After the article was published, Professor Robert Palgrave questioned the quality of the experimental analysis in a series of tweets.

Professor Robert Palgrave claimed that for all five MxSb4-xPb4O13 compounds reported in our work, A-Lab lab experiments resulted in only Sb2Pb2O7, based on the similarity of their diffraction patterns. In this paper, we provide further experimental evidence that the target compounds in our work have indeed been successfully synthesized as mentioned in this paper.

Professor Robert Palgrave also believes that the Rietveld refinements of several compounds we have provided are very poor and contain large residuals. We would like to clarify that the purpose of our work is to show what independent laboratories can achieve, not to show the best results that can be achieved outside the human A-Lab. We agree that self-recycling is challenging in this regard and we look forward to working with the scientific community to further improve automation methods.

In response to the link to the following article, Gerbrand Ceder specifically clarifies the problems pointed out by Professor Palgrave.

Professor Palgrave claimed that for all five MxSb4-xPb4O13 compounds in the paper, A-Lab 's experiment produced only Sb2Pb2O7, based on the similarity of their diffraction patterns.

However, research and analysis show that this is not correct. The author provides two other pieces of information to confirm the successful synthesis of these compounds:

1. For each sample, the author provides EDS data (figure A), indicating that the additional elements (Hf, Zr, Sn, Fe and In) are well incorporated into the final product.

two。 The shift of the peak position in the XRD spectrum measured by the experiment is consistent with that of ion substitution (figure B). Their positions show an obvious trend with the ion radius of each substituted element, and are obviously different from the Sb2Pb2O7 compounds proposed by Professor Palgrave.

Successful synthesis of Hf2Sb2Pb4O13,Zr2Sb2Pb4O13,Sn2Sb2Pb4O13,FeSb3Pb4O13 and InSb3Pb4O13

Furthermore, more accurate characterization can always increase the credibility of any synthetic sample interpretation, but the consistent peak shift in EDS and XRD and the lack of any large number of impurity combinations indicate that the target phase was successfully prepared.

This refutes Professor Palgrave's assertion that only Sb2Pb2O7 is synthesized in each case.

Scanning electron microscope (SEM) images and energy dispersive spectroscopy (EDS) diagrams An obtained from the synthetic products targeting five MxSb4-xPb4O13 compounds showed that all the newly introduced cations (Fe, Hf, In, Sn and Zr) were retained in the samples and were not lost in the synthesis process.

The elements also seem to be uniformly distributed in the particles, with almost no metal-rich particle areas, indicating the formation of materials containing all the precursor elements (M, Sb, Pb, O), which are consistent with the target MxSb4-xPb4O13.

The magnified XRD pattern of targeting five MxSb4-xPb4O13 compounds is on the right side of figure B. the author lists the Shannon ion radius of each newly introduced element (M) in the octahedral environment. The position of the maximum peak in XRD is inversely proportional to the ion radius of the newly introduced element, indicating that it is indeed incorporated into the target structure. Larger ions (such as Hf4+ and Zr4+) cause a significant expansion of the lattice, which is proved by the shift to a lower angle.

Smaller ions (such as Fe3+) cause less expansion, but it is still very different from the known reference phase Sb2Pb2O7. One ion (In3+) seems to deviate from this trend, but only because its concentration in the target structure InSb3Pb4O13 is lower than that of the ions involving M4 + targets (such as Zr2Sb2Pb4O13).

Improved quality of automated Rietveld

Professor Palgrave believes that the Rietveld refinements of several compounds provided are poor and contain a large number of residuals.

Because A-Lab is analyzed in an automatic way, phase recognition is carried out in two steps.

First, after each synthesis step, the ML algorithm performs phase identification and prompts the possible phases in the sample.

Finally, once the sample with the highest phase purity is obtained (not necessarily 100% phase purity, of course), it will be automatically purified according to the phases proposed by ML.

In addition to the two modes that require human intervention, this article provides the results of this automatic program. We have no doubt that manual labor can refine these samples with higher quality.

However, the goal of the study is to show what autonomous laboratories can achieve, not to show the best (or average) results that humans can achieve outside of A-Lab.

Novelty of synthetic compounds: MnAgO2 and Mg3Ni3MnO8

Professor Palgrave pointed out that MnAgO2 and Mg3Ni3MnO8 are not "new" compounds.

The researchers agreed with this assessment.

The structures of these compounds did not exist in the earlier version of ICSD, and we examined them according to ICSD and marked the material project items accordingly, but these two stages have been reported in the literature before.

We do want to point out that A-Lab does not provide information about the synthesis of these targets in its training set derived from the literature.

As a result, these synthetic attempts will still be considered "successful" because the laboratory has successfully synthesized a compound without synthetic formula information. We will certainly update the published papers.

Reference:

Https://twitter.com/Robert_Palgrave/status/1730358675523424344

This article comes from the official account of Wechat: Xin Zhiyuan (ID:AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.