Machine learning and fluid mechanics

Recently Physical Review Fluids published an invited op-ed, "Perspective on machine learning for advancing fluid mechanics" by Brenner, Eldredge, and Freund.  As one who has dabbled in both fields, I found their comments illuminating and balanced, conveying both astonishment and excitement at the achievements and potential of machine learning methodology, as well as its potential pitfalls and limitations.  I share their astonishment that in some cases, deep neural networks have been able to generalize beyond the training data.  This remains an ill-understood phenomenon.  We do not know under what circumstances such generalization can reliably occur, and I believe any such claims about these generalizations must be validated with independent data sets.  Regardless, however, I am in broad agreement with the authors' views, including their conclusion that machine learning methods have potential for high impact "so long as outcomes are held to the long-standing critical standards that should guide studies of flow physics."

Reference

M. P. Brenner, J. D. Eldredge, and J. B. Freund, 2019:  Perspective on machine learning for advancing fluid mecanics.  Physical Review Fluids, 4:  100501 (7 pages).

Science Magazine embraces replication (and gets its nose bloodied)

Science has put its money where its mouth is.  In this week's issue, the editor Jeremy Berg writes about "Replication Challenges".  They have published not one, but three NIH-funded replication studies of an earlier paper.  That is the good news.  The bad news is that a sister journal published a 10-patient human clinical trial, also based on the original paper, in parallel with the 3 replication studies.  Like all 3 replication studies, the human study was not able to reproduce the effect claimed in the original paper.

Can you see why journals may be reluctant to press forward with publishing replication studies?  They can certainly get their noses bloodied.  DTLR commends the editor of Science for restating his commitment to publishing replication studies.

Jerry Gollub, 1944-2019

The first substantive post I wrote for this blog, back in 2013, was a commentary on a paper of Jerry Gollub's, on continuum mechanics in physics education.  I learned recently that sadly, Dr. Gollub lost his life earlier this summer.  An obituary by his renowned collaborator, Harry Swinney, has been posted at Physics Today.  See also this obituary posted at his institution, Haverford College.  I never met or corresponded with Dr. Gollub, but his work has been an influence on mine since I was a junior in college, when I studied from his book with Gregory L. Baker, Chaotic Dynamics:  An Introduction.  In graduate school, the book he edited with Swinney, Hydrodynamic Instabilities and the Transition to Turbulence, was of equal importance to my work.  In my student days, I had these books checked out "long term" from the library, and eventually acquired copies of my own.  DTLR joins the physics and fluid mechanics communities in mourning the loss of this inspiring scientist. 

Since I completed my doctoral thesis in 2001, several other titans I cited in it have passed.  These include V. I. Arnol'd (1937-2010), Robert P. Behringer (1948-2018), Philip G. Drazin (1934-2002), Louis N. Howard (1929-2015), Daniel D. Joseph (1929-2011), Leo P. Kadanoff (1937-2015), E. Lothar Koschmieder (1929-2017), Robert Kraichnan (1928-2008), Olga Ladyzhenskaya (1922-2004), Edward N. Lorenz (1917-2008), John L. Lumley (1930-2015), Steven R. Orszag (1943-2011), William Hill Reid (1926-2016), Philip G. Saffman (1931-2008), James B. Serrin (1926-2012), and Norman J. Zabusky (1929-2018).  There may certainly be others I am not aware of.

References

G. L. Baker and  J. P. Gollub, 1996:  Chaotic Dynamics:  An Introduction.  Second edition.  New York:  Cambridge University Press.  (First edition, 1990.)

H. L. Swinney and J. P. Gollub (eds.), 1985:  Hydrodynamic Instabilities and the Transition to Turbulence.  Second edition.  Topics in Applied Physics, Vol. 45.  Berlin:  Springer.  (First edition, 1981.)

The New Yorker weighs in on statistics

Hannah Fry writes in the New Yorker this month about the benefits and pitfalls of statistics.  Highly recommended.

Reviews of Modern Physics 90th Anniversary

The February issue of Physics Today celebrates the 90th anniversary of Reviews of Modern Physics, the review journal of the American Physical Society (APS) and sibling of the Physical Review series journal DTLR has recently discussed on this blog.

Reviews of Modern Physics (RMP) has occasionally included review papers on fluid mechanics, though in an APS News post, the editor states that he wishes to expand to solicit articles on fluids.  This would be in keeping with the new Physical Review Fluids journal that recently joined the APS stable.  Of course, informed readers will know that the undisputed leading review journal in fluids is the superb Annual Review of Fluid Mechanics, founded 50 years ago (1969) this year.

Physical Review E milestones, concluded; and a look back at the 100th anniversary collection.

With the new year, the list of Physical Review E milestone papers that I previously wrote about appears to have been concluded.  To recap, the editors were posting one paper for each year of the journal's 25 in existence, and I drew attention to those papers with the greatest relation to fluid dynamics.  At the time of my last post, the editors had reached 2011, and now they seem to have concluded with 2017.  The most notable fluid dynamics-related paper in the post-2011 cohort is Settnes & Bruus (2012), "Forces acting on a small particle in an acoustical field in a viscous fluid."

In the meantime, I acquired a massive tome published a quarter century ago, The Physical Review:  The First Hundred Years, A Selection of Seminal Papers and Commentaries, edited by H. Henry Stroke (AIP, 1995).  The book includes a CD-ROM with additional papers not reprinted in the book; this CD-ROM was reissued in 1999 with improved software.  Chapter 6 on Statistical Physics, edited by the late Joel L. Lebowitz, includes work on fluid mechanics.  (Chapter 10 on Plasma Physics, edited by Marshall N. Rosenbluth, has related material.)

Naturally, the 1931 Onsager papers on reciprocal relations in irreversible processes, and Halperin & Nelson (1978) on 2D melting, which I discussed previously, appear here as well.  Also reprinted in the book are Onsager's 1945 famous contribution to turbulence theory, "The distribution of energy in turbulence," one of several independently proposing the famous -5/3 scaling law of Kolmogorov and Obukhov, also claimed by Heisenberg and von Weiszacker.  Furthermore, the classic papers of Ahlers (1974), "Low-temperature studies of the Rayleigh-Benard instability and turbulence", and Gollub & Swinney (1975), "Onset of turbulence in a rotating fluid," are reprinted.  Many of the other papers in this chapter are closely related to fluid dynamics as well, as are many reproduced only on the accompanying CD-ROM.  Of the latter, I will point in particular to Siggia & Zippelius (1982), "Pattern selection in Rayleigh-Benard convection near threshold" and Brandstater, Swift, Swinney, Wolf, Farmer, Jen, & Crutchfield (1983), "Low-dimensional chaos in a hydrodynamic system."

Strangely, the Frisch, Hasslacher, & Pomeau (1986) paper on lattice gas cellular automata, the only fluid dynamics paper noted in the 125th anniversary milestone list that I wrote about earlier, did not appear at all in the earlier 100th anniversary collection.  And of course, none of the landmarks papers from Physical Review E, founded in 1993, and thus too recent to be covered in the 100th anniversary volume, appear there.  In this respect, I find that the 100th anniversary volume, which included literally hundreds of papers overall, provide a somewhat better representation than the far more limited 125th anniversary landmark list.  Still, as I argued in my earlier posts, the history of fluid mechanics research is probably better told through other journals with a specifically elite reputation in fluid mechanics.

Scientific American on "How to Fix Science"

The October 2018 issue of Scientific American features a set of articles under the umbrella, "How to Fix Science."  It begins with a prologue describing "antiscience currents" and arguing that one of the ways to "fight back" is to "tackle its own problems" and "shore up their enterprise from the inside."  The articles in the feature are as follows.

• "Rethink funding" by John. P. A. Ioannidis.
• "Make research reproducible" by Shannon Palus.
• "End harassment" by Clara Moskowitz.
• "Help young scientists" by Rebecca Boyle.
• "Break down silos" by Graham A. J. Worthy and Cherie L. Yestrebsky.

The middle three articles are by science journalists, while the first and last are by academic scientists. Readers of DTLR will recognize that the first two articles discuss topics that have been kicked around on this blog over the years.  DTLR readers will also notice that I have been castigating Dr. Rush Holt, CEO of AAAS, recently, for avoiding these very issues.  Holt believes science is just fine the way it is.  DTLR has previously made a similar argument as the prologue of the Scientific American feature:  we need to clean up our own house before we can, with any good conscience, ask the public to continue financially supporting our endeavors.  Thus, I view the Scientific American piece as a rebuke to the Rush Holt model for responding to the "significant erosion of trust in science over recent years" (a phrase used by Worthy & Yestrebsky).

The leading piece by Ioannidis is a good starting point for this conversation.  Perverse incentives stimulate perverse outcomes.  His piece not only touches on funding issues, but hiring, promotion, and tenure decisions. While I disagree with many of the details in his proposed solutions, I believe the spirit of his piece is absolutely on target.  We need to fix the incentive systems in the scientific enterprise, and this is fundamental infrastructure.  Unfortunately the elites who thrive in the current scientific infrastructure have a vested interest in maintaining the status quo, and would oppose all Ioannidis' proposed reforms.


The article by Palus is a decent introduction to the non-reproducibility issue for the general reader, but lacks the space to delve into the methodological details of good experimental design, experimental quality control, and adequate reporting that are discussed, for instance, in Richard Harris' recent book, Rigor Mortis.

Moskowitz states that sexual harassment is more prevalent in academia than in any other arena except the military.  I am not surprised by this.  Having worked in various sectors of the economy - academia, industry, and government - my personal preference for a setting conducive to scientific research is the private sector, not academia.  In fact, the academic tenure system protects 'assholes' including those who practice harassment.

Boyle's article consists of comments by various young scientists on the following issues:  moving, money, culture, family, industry vs. academia, getting jobs/fellowships/into school, and representation and inequality.  I agree with the writer as well as Ioannidis that the treatment of young scientists in today's science ecosystem is appalling, and needs to be drastically reformed.

Finally, the article on interdisciplinary research is really another reflection of the perverse incentive systems for academic science.

I mentioned that I agreed with the spirit, but not many of the details of Ioannidis' piece.  Readers, what problems do you see in the incentive structure for scientists, and what solutions would you propose?