info@biomedres.us
+1 (502) 904-2126
One Westbrook Corporate Center, Suite 300, Westchester, IL 60154, USA
Site Map
ISSN: 2574 -1241
ISSN: 2574 -1241
Received: June 26, 2026; Published: July 06, 2026
*Corresponding author: L Ridgway Scott, The University of Chicago, Emeritus, USA
DOI: 10.26717/BJSTR.2026.66.010288
We discuss three different technical domains in biomedical research where simulation can be used to populate data deserts, all of which occur at three distinct physical scales. Moreover, we describe significant computational challenges in doing so for more demanding applications. In the case of fluid flow, we give precise Reynolds numbers that demarcate the current boundary between doable simulations and ones that require some advances in technology. The term “data desert” appears in many fields, including medicine [1,2], media research [3], meteorology research [4], as well as in many areas of technology [5]. These are parts of data space that are sparsely populated. This often limits the ability of AI to make reliable predictions [6]. On the other hand, there are many areas in which data is plentiful. Just picking fluid dynamics as an example, important technological phenomena are well studied experimentally [7] and computationally [8,9]. But there are many other areas of fluid mechanics where data is sparse. The purpose of this paper is to explore the use of simulation to populate data deserts.
In many areas, this approach does not work. But in areas in which there is a mathematical model for the phenomena of interest, this approach can potentially be effective. But it is critical that the simulation technology is reliable, which often requires advances in numerical analysis [10], algorithm design [11], and software development [12]. Often, extensive comparison of simulation techniques provides valuable guidance [9]. Moreover, the equations used to model the phenomena may have to be chosen carefully. Reduced models are frequently used which may not be sufficient. Our goal is not to settle all possible questions. Rather it is to raise issues that may lead to more reliable simulations, and thus to pose open questions that can lead to future developments. Undoubtedly, all of this will require continued advances in high performance (parallel) computing [13]. We have chosen three areas of biomedical research at different spatial and temporal scales to indicate how pervasive the issue of data deserts is. All three of these areas have significant impact outside of biomedical research, but for lack of space we do not describe them.
We give a simple example. There is a lot of discussion on the web about what life can be like if you have $1,000,000 in retirement savings. One might ask: how many people have saved $1,000,000 by retirement? Google AI says that there are 497,000 such Americans. But if you ask: how many people have saved $10,000,000 by retirement, the response is murkier. Data is given for $2,000,000 and $3,000,000, but “the number of individuals with $10 million in dedicated retirement accounts is a fraction of that figure” is all that Google AI tells us. We are left to wonder what that fraction might be. One could extrapolate the figures for k million dollars for k = 1, 2, 3 to k = 10, but it is not at all clear what extrapolation formula to use. The number for k = 10 is probably still not small, but so far this number is not in the collective data base. Of course, if it were needed for some reason, one can imagine it could be estimated accurately. But at the moment this is a data desert. It is likely the case that, although we are currently overwhelmed by massive data sets, the known parts of the data universe is probably only a tiny fraction of what could be known. Think of a fractal set of dimension ϵ for ϵ close to zero.
Yes, but not reliably. The example of retirement earnings is typical. If we knew the wealth distribution, we might be able to discern an extrapolation formula. But then there could be the issue of predicting extreme events [14].
Thus we limit discussion here to data interpolation, not data extrapolation. That is, we assume that data is abundant enough to make reliable predictions. When data is sparse, we propose simulation as a way to populate data space. Of course, this is limited to situations where reliable models are available. But we now consider three such cases.
Protein interactions provide a basis for drug discovery [15]. One of the greatest achievements of AI was recognized by a Nobel Prize in 2024 for protein folding. This accomplishment is based on vast data bases of protein structure solutions. These data bases have been used extensively to make predictions of protein behavior in many areas [16]. On the other hand, these data bases are sparse for portions of proteins which are not well structured, dubbed plastic regions in [5]. It was proposed in [5] that simulations could be used to predict possible structures that could be precipitated by a binding event. For small plastic regions, there is reason to expect that this approach could be successful. In [5], several simulation methodologies were suggested, but there are models at different levels of fidelity. At the highest level, quantum mechanics would be used, but this may be computationally prohibitive even for a few residues. Moreover, even classical quantum mechanics is being questioned in some circles [17]. An intermediate level would be Quantum Mechanics/Molecular Mechanics (QM/MM) methods, which were recognized by a Nobel Prize in 2013. Standard molecular modeling (MM) assumes properties of molecular interactions that may be inaccurate (incorrect physics) in some situations. The competition, namely experimental structure determination [6], uses real physics, not approximations. That is, the experiments determine behavior that may be difficult to simulate.
It is possible to compute the solution of quantum mechanics equations for small systems accurately [18-20]. But at the moment, such technology has not yet been extended to larger molecular systems.
Fluid mechanics governs many medical issues in the human body. Cardiovascular disease is perhaps the most well known [21,22], but fluid flow in the brain [23] is also a very active area of research, as well as in other organs [24]. Everyone who has flown on a commercial aircraft is aware of the word turbulence. This word is used to explain the bumpy ride that often occurs. But it turns out that turbulence is not a precisely defined concept. We know it when we see it, but there are many different fluid flows that could be described as turbulent. One such event occurs with flow in arteries [21,22]. Since commercial and private aircraft fly in large numbers everyday, it may seem counter-intuitive that there is a data desert here. But the problem is that instrumenting a commercial aircraft to measure the relevant data would be complicated. Unlike the case of molecular modeling in the previous section, the model equations for fluid flow are well established: the Navier–Stokes equations [25]. On the other hand, solving these equations accurately is challenging. Reduced equations, such as Reynolds-Averaged Navier–Stokes equations (RANS) [26] are popular, but inaccurate. Simulation of fluid flow has advanced substantially in the past fifty years. We now know how to identify reliable simulation technologies for fluid flow, but many important concepts are still in development.
For example, a basic benchmark in fluid dynamics is flow past a cylinder. For low speed flows (low Reynolds numbers), this can be done with sufficiently sophisticated techniques [9,27] for two-dimensional flows, but this benchmark shows the flaws in many simulation algorithms. For three-dimensional flows, the state of the art is more rudimentary. Thus it may take some time to fully populate the turbulence data desert. To provide a more complete picture, we focus on flow past a cylinder to give a sense of where simulation technology stands today. More precisely, we begin with a cylinder of infinite length and uniform flow (at infinity, or at a large distance upstream from the cylinder). We assume the flow direction is perpendicular to the cylinder axis. We will describe how the character of the flow changes as the speed is increased, but for precision we will use the Reynolds number as a descriptor for flow speed. The Reynolds number R is defined by R = UL/ν where L is the cylinder diameter, U is the flow speed at infinity, and ν is the fluid kinematic viscosity.
For R less than about 50, the flow is steady, and up-down symmetric, where we think of the plane perpendicular to the cylinder axis as the x, y plane, with z denoting the third dimension along the cylinder axis. More precisely, the velocity u satisfies u(x, y, z, t) = (u(x, y, z), v(x, y, z), 0) with symmetry u(x, y, z) = u(x, −y, z) and v(x, y, z) = −v(x, −y, z). Around R = 50, the flow becomes unsteady, but still planar (the component of u in the z-direction is zero). What is observed is the Karman vortex street [28], where vortices in the x, y plane are generated alternately from the top and bottom of the cylinder. The drag on the cylinder has been measured extensively experimentally [7] as well as computed via simulations. One measure of the flow is the drag-lift phase diagram [29]. This phase diagram remains periodic up to R = 1000, but it becomes chaotic by R = 2000 [9,30]. But notably, many simulations get this wrong, including [29].
Simulations in two dimensions are of significant industrial value for R up to a few thousand [31], but by R = 10, 000 significant three-dimensional effects appear [32]. Beyond that, planar flow breaks down. It is suggested [33] that at some point the two dimensional vortices turn ninety degrees and instead a new type of vortex appears whose axis is parallel with the flow direction, instead of parallel with the cylinder axis as occurs for low Reynolds numbers. This has not been independently validated but it makes sense; if the lift oscillation continued to grow [34] then many cylinders (light and telephone poles, for example) would fall down. If the vortices turn 90 degrees, lift oscillation is essentially eliminated. To summarize the state-of-the-art, we are now able to reliably compute flow past a cylinder for Reynolds numbers up to a few thousand. On the other hand, many older simulation schemes break down even before R = 1000. This means that doing reliable simulations for R > 10, 000 are a current challenge. But twenty years ago, reliable simulations for R = 1000 were not available, so we can project that advances in algorithms, software, and hardware will allow simulations for significantly higher Reynolds numbers.
Molecular modeling using quantum mechanics (Schrodinger’s equation) reached a pinnacle with the 1998 Nobel Prize awarded in part for the development of the Gaussian code. That code has been widely used to predict structures of bonded molecular interactions. But nonbonded interactions, between molecules that are separated by several Angstroms, are more difficult since the approximation of functions in Gaussian are Gaussians, which decay too fast. Nonbonded interactions are critical components of molecular modeling codes [35-37]. Yet their approximation is often ad hoc. It is possible to approximate some nonbonded interactions (van der Waals forces) by solving the high-dimensional Schrodinger equation [18,20]. But so far, this has been carried out only for very simple systems (two hydrogen atoms).
We have indicated three different technical domains where simulation can be used to populate data deserts. Moreover, we have described significant challenges in doing so for more demanding applications. In the case of fluid flow, we have given precise Reynolds numbers that demarcate the current boundary between doable simulations and ones that require some advances in technology.
Abstract
PDF
