A recent Twitter thread got me thinking back to the frustrations I felt working towards gaining non-research employment during the last year of my physics PhD program at Columbia University from 2014-2015. While I tried to participate in the thread, I had way too many thoughts for Twitter. A blog post felt more appropriate, so here it is.
In summary, the thread was about the failure of physics PhD programs to provide adequate support for students who go on to non-research careers (i.e. a career which is not a post doctoral/research scientist position at a university, national laboratory, or physics-y company like IBM). I wholeheartedly agree that a failure exists, and I would like to use this post to explicate on my very subjective answers to the following questions:
- Why does this failure exist?
- What is the point of a physics PhD program?
- What can be done to improve things?
Full disclaimer: all of my (probably cynical) opinions here are from me generalizing about my own, specific experiences in grad school and from observations of my colleagues. I am sure that there are schools that do things better than Columbia, as well as schools that do things worse.
Why does this failure exist?
While most students intend to pursue a research career on entering graduate school, many do not actually end up in such a career upon graduating. I would argue that one should avoid a physics PhD altogether if they do not initially intend to pursue a research career. Nevertheless, clearly many students change course during graduate school. The crux of the failure is that there is a fundamental misalignment of incentives between students pursuing research careers versus those who are not. The entire system is setup to incentive research careers thus performing a great disservice to all those who change course.
To understand this, let’s start with the incentives for research careers. As with most things, we follow the money. Scientists are perpetually beholden to large organizations for grant money, and the most valuable assets to one’s grant applications are authored, published, scientific papers in prestigous journals. We can consider papers to be the “currency” of academia, and, the more prestigous the journal and higher your name is up the author list, the more valuable that currency is to you. As such, it is and should be every PhD student’s goal to rack up as many high-value papers as possible.
There are many other aspects of academia in which a student may participate, such as building scientific instrumentation, developing software packages, education and teaching, etc… From my point of view as a PhD student, these were all orders of magnitude less important than papers when applying for: post doctoral fellowships, professorships, grant applications, and so on. While this may not actually be the case, this was how I and others perceived things, and so we oriented our efforts accordingly.
To top this all off, the acadmic job market is extremely cuthroat (there is a reason so many students drop out of the professorship pipeline!). As such, it feels as though one most go “all in” on a research career. This means working insane hours for already small pay (e.g. for me, $30K/year for 6 years in New York City). This also means that there are many activities like those that I listed above (e.g. building scientific instrumentation) that are not in a student’s immediate best interest.
Once a student has indicated to a professor that they are no longer interested in a research career, their goals diverge from the professor’s. Speaking from the data science industry, the starkest difference is that literally nobody cares what papers you have published. For the professor, they lose out because the student may no longer work insane hours (effectively becoming more expensive), their interests may stray (e.g. towards software engineering), they will not be a future scientific collaborator, and they will no longer represent good marketing for prospective graduate students (prospective students and post docs like to see that professors’ alumni have gone on to professorships themselves). Whether explicitly or implicitly, the professor-student relationship necessarily changes.
From the student’s point of view, all of the above leads to significant downsides to “showing one’s cards” and revealing a non-research preference. This could be mitigated by potential upsides, such as professors providing guidance towards non-academic career paths, shifting a student’s research work towards more industry-specific skillsets, etc… but I have never seen this happen. There are basically a whole bunch of cons and few, unlikely pros.
What is the point of a physics PhD program?
One could consider my above complaints and argue that the point of a physics PhD program is to train students for research careers and this it is out of scope to expect programs to provide additional support. However, it remains the case that there are simply not enough research positions compared to PhD students. The Academy is then left with three options:
- Continue to ignore this inconvenient truth and do a disservice to non-research career students.
- Accept fewer PhD students such that the professorship pipeline input size better matches the output.
- Invest in better servicing non-research career students.
I would argue that the first option is simply unfair, especially when students are providing such low-wage, highly skilled work. The second option seems unlikely because one needs the existing numbers of students to maintain research labs. Perhaps research scientists and/or technicians could fill this need, though.
I am clearly in favor of the last option. In fact, it seems like undergraduate programs figured this out a while ago. Obviously, not all philosophy or art history majors get jobs in their respective majors. Instead, universities and students decided that a liberal arts education that provides a well-rounded scholarly exposure combined with deep mastery in a particular major is a fair trade for a tuition fee (or at least was a fair trade!). For learning and focusing only on specific skills, we have technical schools. We could make physics PhD programs more aligned with technical schools, but I think this is the wrong approach.
So what would a liberal arts physics PhD program look like? I think back to what experiences, competencies, and values I truly appreciate from my time in graduate school, and there is an overarching theme of building mathematical and technical confidence.
Graduate physics courses are really hard. Some are unnecessarily hard (Jackson’s E&M book is hard enough. Not sure why my professor made us do everything in an arbitrary number of dimensions…). Along the way, you learn advanced mathematics and generally streeeetch your brain in ways that you did not know possible. While this process builds a solid foundation in physics, the true value lies in growing the confidence within the student to be able to tackle any complex mathematical topic that is thrown their way. Sure, physicists get a bad rap for blindly thinking they can solve problems well outside their domain, but there is also a reason physicists are so successful when they do commit to other domains of study.
This confidence was helpful in my physics research. If I could learn Quantum Field Theory, then it was not too much of a leap to think that I could work through the equations pertinent to my specific area of research. From the industry side, I remember being surprised to learn that Deep Learning was “just” multivariate calculus and some signal processing.
As an experimentalist in a new professor’s lab, I primarily spent my time
- Desigining scientific instrumentation in 3D CAD programs
- Machining components of instrumentation
- Assembling instrumentation
- Eternally debugging instrumentation
- Running experiments and collecting data
- Analyzing the data in MATLAB
- Numerical simulations in MATLAB
- Writing papers
I spent more time on the bullet points higher in the list during the beginning of my PhD as the lab was being built. Along the way, I ended up gaining reasonably deep competancy in a couple technical areas, like CAD programs, metal machining, and computational linear algebra. I gained these competencies because these were tools that I needed to solve research problems.
With each tool that I learned, I added another data point of support to my hypothesis that, if I need to learn some deeply technical technique in the future to solve a problem, I will likely be able to teach myself and be successful. Each data point has increased my confidence, and the context around the data points further increases that confidence. For example, there is nothing quite like building a custom $X00,000 scientific instrument and having it break. There is nobody to call to fix it. You will find a way to fix it because you have to fix it.
While I’ve been talking about things in the weeds like classes and machining, the whole point of Physics is to learn more about our universe and actually do Science. This involves things like running experiments or calculations, interrogating the data and your hypotheses from every angle, making a case for your theory or finding, and communicating via papers and talks. This process takes a long time because research deals with the unknown. Perhaps one collects data that looks interesting and then has to figure out what it actually means (that was my group’s MO). Or, it may take years to design and build an experiment to test a hypothesis (see particle physics, dark matter experiments, etc…).
Banging one’s head against the wall trying to figure out something which literally nobody else in the world knows is a valuable exercise. Not everything gets figured out, but some things do. And again, this builds confidence that it’s possible to figure things out and to know the previously unknown. Of course, we’re never really done. Answers bring new questions, and this iterative cycle breeds curiosity. Curiosity leads to persistence and perseverence which have proven immensely useful for me after graduate school.
Lastly, being able to distill a mess of experiments and/or calculations into a well-crafted paper of figures, equations, and text is extremely useful, as well as the ability to communicate complex phenomena in a simple way. No matter your industry or field, no “higher up” wants to see your raw data, notes, code, or numbers.
What can be done to improve things?
Given my cynicism in the first section and the values of a PhD in the second, I’ll close with a look at how a PhD program can help both research and non-research career students.
Columbia may be uniquely poor in this aspect, but I received zero guidance following the passing of my qualifying exams during my first year of graduate school. There was no thesis proposal, no thesis committee (except for the professors assembled a month before my defense), nonexistent requirements for graduation beyond X credits of courses, and no real range of program duration (I’ve observed everything from 4 - 10 years). It’s like I floated for 5 years, spent another year asking to please get out of the water, and then they finally acquiesced.
Early and frequent career guidance would be hugely beneficial for students. For students who are interested in the academic track, a guidance counselor to keep them focused would do wonders. I’ve seen students spend 4 years building a complex scientific instrument only to be told that they still have a ways to go to graduate because they have not taken any data let alone published a paper. Conversely, I’ve seen students who are 100% sure that they want to go into industry and are still trying to push out more papers because they think companies care when in reality they could be pursuing other projects that are still beneficial for research but align better with their career (e.g. refactoring a group’s shared software library).
I am pessimistic that we can rely on all professors to be this career counselor. If nothing else, many professors are wildly out of touch with industry needs (see the prevalence of FORTRAN). As well, there is the fundamental conflict of interest between a professor’s top priorities and a non-research career student’s. Perhaps an unbiased counselor familiar with the domain or even paid physics alumni could be used.
Freedom to change course
While it’s clear that students change from pursuing a research career to a non-research career, the converse can happen, and things can flip flop back and forth. Students need the freedom to try out different things in a low risk way. I have a close colleague from grad school who was seriously considering a career far from science and technology many years into his PhD program due to a lack of research/publishing success. He eventually struck gold and published multiple, excellent papers very late into his PhD and is now well on his way to a tenure-track professorship. We would not want guidance counselors to heavily steer people like him away from a research career.
How do we allow students the freedom to experiment? One way is to require students to maintain a diversified “portfolio” of graduate school experience. Just as some programs have students undergo research rotations to expose them to the range of research options prior to their picking of a research lab, physics PhD programs could require students to obtain a diversity of skills or competencies in order to graduate. I previously mentioned some of the skills that I learned during my PhD. Perhaps pure theorists would be required to use a computer program for some project. Perhaps computuational physicists would be required to use both simulations and machine learning to solve research problems. Maybe this would encourage cross-disciplinary research or at least collaboration between groups because the student would need to seek out experts on these techniques who may lie outsde their group.
Requiring students to do this absolves them of responsibility for time spent outside of their professor’s preferred use of their time. The downside is that this system is likely to be gamed (how do you actually measure these things?). Additionally, students may feel that this program puts them at a disadvantage compared to other programs that don’t have this requirement due to the very fact that the student can’t go “all in” on pure, paper-focused research.
Adjacent to requiring a diversity of competencies, the freedom and encouragement (and possibly even requirement) to do an internship in industry for a summer could also serve as a lightweight way to try out non-academia. What’s odd is that this is fairly common practice in Engineering PhDs while almost anathemna in Physics. I would argue that the 2-3 months of lost research time for such an internship could easily be recouped in the increased productivity of the student after the internship. Why would they be more productive? Personally, I can attest to my own dramatic improvements in time management, email punctuality, software engineering, and more after having spent a couple weeks at my first job. Even learning how to use Google Calendar to schedule meetings would have saved a bunch of time (vs. a 40-email chain with 3 professors in order to nail down a meeting time).
While I would not have wanted to take any more graduate courses than I had to, I do think that courses in both statistics and software engineering would have been hugely helpful. Obviously, these courses would have helped to better prepare me for a career in data science, but I think they would have paid dividends both during my PhD and if I had stayed in academia. For a field known for its high precision, six sigma measurements, I’m amazed that I never had to take a statistics class during either undergraduate or graduate school. Weirdly, physics students end up learning a bastardized version of statistics when they start taking quantum mechanics and statistical mechanics. A proper probability and statistics course prior to those physics classes would have made learning the physics material much easier (seriously, learning about probability density functions is complicated enough, so combining that with infinite-dimensional wavefunctions is just sadistic). As well, a solid foundation in statistics is broadly useful across most technical careers.
And what’s a now-essential part of statistics? Programming. I was lucky enough that my undergraduate physics program required a numerical algorithms course, which was the only programming class that I ever took. Of course, this class was tought in the math department, and it was the luck of the draw if you got the class taught in MATLAB or the class taught in FORTRAN (this was in 2007, but still…). Learning a modern software language along with engineering best practices would have saved me significant time in graduate school. How might a course like this work? I’m imagining a class that covers some of the primary programming techniques that are useful across physics research:
- Regression (including Machine Learning)
- Numerical solutions (e.g. solving laplace’s equation numerically on a grid)
- Possibly data acquisition, manipulation, and storage
Along the way, the student learns how to use Python, GitHub, write packages, unit tests, etc…
So, my reaction to the original Twitter thread ended up in a love/hate letter to Physics Graduate School. I would love to hear about your experiences from other programs, whether the experiences are about improved programs or corraboratory grievances. Please reach out at email@example.com or @eprosenthal.