Battery R&D programs have more compute today than they did ten years ago. Exascale HPC systems are becoming accessible. Cloud computing has reduced the barrier to running parallel DFT jobs. MLIP inference is faster by orders of magnitude than it was in 2018. And yet: the time from a new cathode composition hypothesis to a confirmed synthesis candidate is still, for most programs, 12–18 months. Sometimes longer.
This is not primarily a compute throughput problem. It is a workflow fragmentation problem. Understanding the distinction matters for anyone trying to fix it.
The anatomy of an 18-month discovery cycle
Walk through the actual sequence of a cathode discovery program that ends in a credible synthesis candidate:
Month 1–2: The research team identifies a hypothesis composition. A graduate student or postdoc sets up DFT calculations — or tries to. Setup involves: choosing appropriate pseudopotentials, determining the correct Hubbard U parameters for the transition metal species, converging the k-point grid for the unit cell, constructing a supercell for the substituted composition, and deciding which symmetry-inequivalent arrangements to calculate. For a team without a dedicated computational scientist, this step alone can take weeks. For a team with one, it takes days — but that person is a bottleneck shared across multiple projects.
Month 2–4: DFT calculations run. Each structure takes 2–8 hours. A systematic scan of 50 compositions in a 24-atom supercell runs to ~2,000 CPU-hours. On a typical university HPC allocation, that is weeks of wall-clock time, interspersed with queue waits and failed jobs that have to be restarted. The result, if things go well, is a set of formation energies and a rough stability ranking.
Month 4–5: The computational results are analyzed. Formation energies are referenced against a convex hull — but which hull? The Materials Project hull for the same chemical system? A custom hull built from the same DFT settings? The choice of hull matters by 10–50 meV/atom, enough to change the ranking. This decision is rarely documented in the program's records.
Month 5–7: The top 5–10 computational candidates are forwarded to synthesis. A synthetic chemist sets up solid-state reactions or co-precipitation. First synthesis runs take 2–4 weeks. Characterization (XRD, ICP, electrochemistry) takes another 2–3 weeks. Some candidates fail at synthesis — wrong phase, impurity phases, cation disorder too high. The information from synthesis failure is rarely fed back into the computational model to understand why.
Month 7–12: Surviving synthesis candidates are cycled. Electrochemical characterization reveals capacity, voltage profile, rate capability. Some candidates that looked excellent by DFT formation energy perform poorly in cells — because the DFT screening did not include surface energy, phonon stability, or migration barrier calculations that would have predicted the failure mode. The program pivots, generates new hypotheses, and restarts partial loops of the above.
Month 12–18: With luck, one or two candidates survive to serious consideration. The program has consumed a PhD student-year or more of compute and wet-lab time per viable candidate.
The actual bottlenecks
Looking at this timeline, three bottlenecks stand out — none of them are "not enough CPU hours."
Bottleneck 1: Setup expertise. DFT calculations for transition metal oxides require specific methodological choices that significantly affect the result. Using the wrong U parameter for Ni can shift the predicted formation energy by 100–200 meV/atom. Using a too-sparse k-point grid introduces errors of similar magnitude. Using GGA instead of GGA+U for Mn-containing compounds produces qualitatively wrong electronic ground states. These are not academic concerns — they cause screening campaigns to produce unreliable rankings that correlate poorly with synthesis outcomes. The knowledge required to set up DFT correctly for this chemistry is held by specialists, and most battery R&D programs have fewer of those than they have chemists and electrochemists.
Bottleneck 2: Workflow fragmentation. The steps described above — enumeration, DFT calculation, hull construction, surface energy, phonon analysis, output packaging — are performed in disconnected tools (VESTA for visualization, a custom Python script for enumeration, VASP for DFT, phonopy for phonons, a spreadsheet for results). There is no standard interface. Each hand-off between steps is an opportunity for error and a time sink. The researcher who set up the DFT jobs is often not the researcher who analyzes the hull, who is often not the synthetic chemist who decides which candidates to make. Information is lost at each hand-off.
Bottleneck 3: Absence of pre-synthesis dynamic stability filtering. Formation energy screening produces a list of thermodynamically stable phases. It does not tell you which of those phases are dynamically stable — which will actually adopt the target crystal structure under synthesis conditions rather than distorting to a different polymorph. Phonon dispersion calculations, which provide this information, are routinely skipped in screening workflows because they require an additional software stack (phonopy, interfaced to VASP), an additional expertise level, and additional compute. The result: candidates that look excellent computationally fail at synthesis because they were dynamically unstable, and no one computed the phonons to catch this before committing experimental resources.
What faster hardware does not fix
Exascale computing and cloud DFT do not address any of these three bottlenecks. A researcher who doesn't know the right U value for Mn will compute the wrong result faster. A fragmented workflow becomes a faster fragmented workflow. The phonon calculation step remains skipped if it isn't integrated into the standard pipeline.
The 18-month problem is primarily a methods integration problem. The solution is not raw compute — it is a workflow where structure enumeration, MLIP pre-screening, DFT validation, phonon filtering, surface energy, and output packaging are connected in a single computational environment, with appropriate method defaults for the target chemistry class, and output in a format that experimental teams can act on directly.
The realistic ceiling of computational acceleration
Computational screening is not going to reduce a cathode discovery program from 18 months to 2 months. The experimental steps — synthesis, characterization, electrochemical cycling, failure mode analysis — have irreducible timelines that computation does not affect. A synthesis run takes the time it takes. An electrochemical cycling study requires the number of cycles it requires.
What computational screening changes is the hypothesis quality entering the experimental program. Instead of 50 compositions that are a rough scan of composition space, the experimental team works with 5–10 candidates that have been pre-qualified on thermodynamic stability, dynamic stability, surface energy profile, and voltage prediction. The failure rate in synthesis and early characterization drops; the fraction of candidates that reach serious electrochemical evaluation increases. An 18-month program for 1 viable candidate becomes a program that produces 3–5 viable candidates in the same time window — because the experimental resources are concentrated on pre-qualified hypotheses rather than spread across an undifferentiated exploration.
This is the realistic scope of what MaterSynq does, and what we tell every lab before they start a campaign: we accelerate the pre-synthesis qualification step, not the experiment. The experiment is still yours to run.